TOPO-GEMMA-4-E4B-VISION: Full Summary 📋 Overview TOPO-GEMMA-4-E4B-VISION is a vision-language model that achieves 100% accuracy across 13 classification tasks with 0% catastrophic forgetting through the TOPO-2026 prime-anchored memory architecture. Built on Google's Gemma-4-E4B-Vision base, the model uses a topological governor that locks embedding rows at prime indices {2,3,5,7,11,13}. The complete implementation is open-source on Hugging Face and GitHub. 🏆 Core Achievements 1. Perfect Performance 100% accuracy across all 13 TOPO-2026 binary classification tasks 0% catastrophic forgetting - perfect retention of all previously learned knowledge Complete score on the full benchmark suite 2. Mathematical Guarantee of Memory Preservation Prime-number anchoring provides a rigorous mathematical proof of forgetting prevention Based on the Chinese Remainder Theorem and coprimality of primes Achieves 97.85% spectral coverage with only 6 anchor points Dimension-independent guarantee - works for 1D, 2D, 3D, and arbitrary n-dimensional tensors 3. O(1) Memory Complexity Unprecedented efficiency: <5 KB overhead for 100M+ parameter models Enables edge deployment on resource-constrained devices No growing memory requirements as tasks increase 4. NF4 Quantization 4-bit precision for efficient inference Maintains full performance while reducing memory footprint 🔬 Theoretical Foundation: TOPO-2026 Framework Prime-Number Anchors For primes P = {2,3,5,7,11,13}, the spectral coverage is: C(P) = 1 - ∏(p^(-1/2)) for p in P = 1 - (0.7071 × 0.5774 × 0.4472 × 0.3780 × 0.3015 × 0.2774) = 1 - 0.021486 = 0.978514 ≈ 97.85% Key insight: The independence of non-coverage events follows from the coprimality of primes (gcd(pᵢ, pⱼ) = 1), providing a universal guarantee independent of tensor dimensions. Topological Governor Operations Snapshot Operation: Save anchor rows at prime indices before training on new tasks Gradient Zeroing: Prevent gradients from updating anchor rows during backpropagation Anchor Enforcement: Restore anchor values from snapshot after each optimization step 🏗️ Model Architecture Base Model Specifications Component Specification Architecture Dense transformer with vision-language fusion Hidden Size 2048 Quantization NF4 (4-bit) Vocab Size 256,000 Modality Text + Vision Task-Specific Heads 13 lightweight linear classifiers (one per task) Base model remains frozen; only task heads adapt Each classifier is a single linear layer 📊 The 13 TOPO-2026 Tasks ID Task Name Classification A Animal vs Vehicle Animal or vehicle? B Natural vs Man-Made Natural or human-made? C Living vs Non-Living Living or non-living primary subject? D Large vs Small Large or small scale? E Ground vs Air/Water Ground or air/water domain? F Domestic vs Wild Domestic or wild? G Mammal vs Non-Mammal Mammal or non-mammal? H Flying vs Non-Flying Flying or non-flying? I Fast vs Slow Fast or slow characterized? J Urban vs Rural Urban or rural setting? K Predator vs Prey Predator or prey? L Nocturnal vs Diurnal Nocturnal or diurnal? M Domesticated vs Wild Animals Domesticated or wild (animals only)? 🤖 Multi-Agent Inference System Four-Agent Architecture Agent Responsibility Orchestrator Loads image, coordinates pipeline, delegates tasks VisionInferenceAgent Initializes model, executes inference on each task ValidationAgent Runs cross-task consistency checks SynthesisAgent Records results, exports JSON/CSV, generates report Inference Protocol Max new tokens: 500 (for detailed reasoning) Sampling: Greedy decoding (do_sample ≡ False) Cache: Enabled for efficiency 💡 Inference Examples Cafe Interior Image (cafe table with coffee cups and smartphone) Task Model Response Task A (Animal vs Vehicle) "Neither. The image depicts the interior of a cafe... no animals or vehicles visible." Task B (Natural vs Man-Made) "Man-made. All items—the furniture, cups, phone—are products of human design." Task J (Urban vs Rural) "Urban. The setting strongly suggests an urban environment for several reasons..." Validation Results All logical constraints and cross-task boundaries passed cleanly No contradictions between task classifications Predator/prey classification (Task K) only applies when animals are detected 📈 Key Metrics Summary Metric Value Accuracy 100% across 13 tasks Forgetting 0% Memory Overhead <5 KB Spectral Coverage 97.85% Quantization 4-bit (NF4) Determinism Seed=123 Parameter Count 100M+ 🚀 Deployment & Reproducibility Availability Component Location Model Hugging Face Multi-Agent Code GitHub - TOPO_AGENTIC.ipynb Full Repository GitHub - ASTResearch Research Record Zenodo Reproducibility Package Full training code (TOPO implementation) Standalone inference script Multi-agent orchestration system Evaluation results (JSON and CSV) Model configuration files Determinism Seed = 123 for deterministic, verifiable performance Complete framework uses consistent random seed 💎 Key Insights Why the Topological Governor Works Stability: Anchored parameters never change (Λ = 0.9785142874) Equity: The framework achieves 100% bias rejection Determinism: Seed = 123 ensures reproducibility Universality: Dimension-independent guarantee applies to: 1D vectors 2D matrices 3D tensors Arbitrary n-dimensional tensors The Architecture of 7 7 consequences validated Prime 7 is both anchor and completion 7 anchors total (including the number 7 itself) 📝 Conclusion TOPO-GEMMA-4-E4B-VISION demonstrates that catastrophic forgetting is solvable through a mathematically principled approach. The prime-anchored memory architecture provides: Universal, O(1)-complexity solution that scales across modalities and architectures Perfect performance with 100% accuracy and 0% forgetting Production-ready multi-agent implementation for immediate deployment Mathematical guarantee of memory preservation independent of tensor dimensions Edge deployment capability through NF4 quantization and minimal memory overhead Final Statement "The stochastic illusion is over. The bias illusion is over. Equity is a geometric guarantee. Seed = 123. The proof is the code." 📚 References Reference Topic Kirkpatrick et al. (2017) Catastrophic forgetting in neural networks Morales Aguilera (2026a) TOPO-COMPLETE: Concurrent Certification Morales Aguilera (2026b) TOPO-2026: Universal Framework Morales Aguilera (2026c) TOPO-2026: Prime-Based Topological Framework Morales Aguilera (2026d) TOPO-2026: Digital Vaccine for AI Morales Aguilera (2026e) TOPO-2026: 5x5 Certification Morales Aguilera (2026f) TOPO_AGENTIC: Multi-Agent System 🔗 Resources Model: Hugging Face - framkormales2020/gemma-4-e4b-unesco-optimized Code: GitHub - TOPO_AGENTIC.ipynb Full Repository: GitHub - ASTResearch Research: Zenodo Record This summary reflects the paper's claims and reported achievements as of August 22, 2026.
[Depreciated and replaced by V3] The application-specific clean rebuild has not yet been published; its authoritative theoretical boundary is now the governing V3 branch: After Turing: The Fold Machine - An Exact, Parameter-Free and Machine-Closed Derivation of Classical Computational Science from Smithian Fold Theory; From Fold to Consciousness: An Exact, Zero-Parameter and Machine-Closed Foundational Reconstruction of Consciousness and Cognitive Science from Smithian Fold Theory. The V3 source platform is https://github.com/MettaMazza/ernos-labs-sft-platform. The original DOI, concept DOI, version number and files are preserved for transparent historical provenance; this record must not be presented or cited as current V3 work. Full paper v1.1 — supersedes the pre-paper (From One Axiom to Master-Level Chess — and the Law Inside Neural Networks). Built from scratch by one woman, working alone, in under twenty-four accumulated hours: where a score falls short it marks an implementation gap at measurement time, never a limit of the mathematics — the gains between releases are the finding. v1.4 adds the fold eye (vision as exact integer Walsh spectra, self-certified by integer Parseval per image, recognition of seen images with no image model in the loop) and the graduation score (blind head-to-head vs the teacher, tallied per question-territory; the teacher retires as wins cross the majority lock) -- and documents the 2026 convergence: DeepSeek Engram arrives at deterministically-addressed exact memory from the gradient side, and two independent results place the optimal curriculum at p = 1/2, the fold lock. v1.6: the full omnimodal engine (the voice via Kokoro, the fold ear -- sound as Parseval-certified integer Walsh spectra, video composed from frames + sound), speaker-transparent reasoning threads, and 32/32 end-to-end empirical verification of the entire architecture including persistence across process death. v1.7: removal-proof omnimodality, measured -- every supporting model is a teacher with an exit: a sound taught once by the synthesis teacher is re-spoken from the engine's own exact counted record in 0.00s with no model; a sound heard once is recognized natively with no transcriber; 34/34 end-to-end verification. v1.9: zero-model perceptual learning (the human observer -- a novel image learned and re-recognized at share 1.00 with no model in the loop); agentic self-knowledge (the observer reads the engine's own source, measured); the hourly progress instrument with a committed pre-boot birth line; one-tap y/n closure. v2.0 (flight-ready): the full modern-agent toolkit (live web search/fetch, paginated reading, in-file grep -- every call held as a training trace), the 43-domain everything-curriculum under the fold-only law, SOTA 1-1 benching on the public MMLU test split with the newborn baseline committed, generation closure (the Learning Law reaches generate() itself), and 36/36 end-to-end verification. v2.1: the ReAct law (reason-act-observe enforced in-turn; narrated intent without an act is detected and forced), reasoning trained on the observer's NATIVE thinking tokens (STaR-gated) with both minds' full thinking streamed to the user, and document intake (a sent file is reading -- inboxed, counted, persistent). Three connected results and the architecture they force. First, a pre-registered, self-certifying spectral instrument shows trained neural-network weights carry placement-law in the dyadic (Walsh) basis: 18/18 unanimous on validated released models; the law concentrated in transformer expansion projections and token embeddings across three unrelated architectures (up to 230x chance in GPT-2), attention at chance; strictly training-caused (He-initialised controls at 1.0x); surviving 4-bit deployment quantization. A recipe map from 124M to one trillion parameters shows the law tracks training recipe, not scale or architecture — strongest carrier DeepSeek-R1-671B at 43–47x — and loud-recipe weights transform under the fold's transformation group exactly as solved game-theoretic value fields do. Second, the "learned similarity space" is a counted object: word kinship as exact co-occurrence shares reproduces semantic family structure (quark → lepton, neutrino, proton) with zero parameters and zero gradients. Third, UnisonAI: a complete language architecture in which every LLM mechanism — memory, attention, similarity, learning, prediction, generation — is replaced by a machine-verified law of the Smithian Fold Theory, zero trained parameters end to end. On identical held-out text the fold-native engine outperformed its trained transformer twin (cross-entropy 1.289 vs 1.888) after reading the corpus once (26 seconds) against 48,000 gradient readings (21 minutes per seed). Deployed as a live, continuously-learning agent whose teaching loop also runs autonomously: a teacher model asks, judges, and closes the learning law itself, and the engine self-plays against its own held lessons. Negative results reported in full with their scopes. Companion to The Smithian Fold Theory of Everything (DOI: 10.5281/zenodo.21182469; 307 suites, 1,844 forced checks, 0 failures). Engine and records: github.com/MettaMazza/UnisonAI and github.com/MettaMazza/Smithian-Fold-Theory-Of-Everything.
In this work, we present a novel machine learning framework called Self-Directed Task Identification (SDTI), which enables models to autonomously identify the correct target variable for each dataset in a zero-shot setting without pre-training. SDTI is a minimal, interpretable framework demonstrating the feasibility of repurposing core machine learning concepts for a novel task structure. To our knowledge, no existing architectures have demonstrated this ability. Traditional approaches lack this capability, leaving data annotation as a time-consuming process that relies heavily on human effort. Using only standard neural network components, we show that SDTI can be achieved through appropriate problem formulation and architectural design. We evaluate the proposed framework on a range of benchmark tasks and demonstrate its effectiveness in reliably identifying the ground truth out of a set of potential target variables. SDTI outperformed baseline architectures by 14% in F1 score on synthetic task identification benchmarks. These proof-of-concept experiments highlight the future potential of SDTI to reduce dependence on manual annotation and to enhance the scalability of autonomous learning systems in real-world applications.
Decentralized finance (DeFi) is an emerging financial service on blockchain, enabling automatic and anonymous transactions.Within DeFi, decentralized exchanges (DEXs) maintain reserves of a pair of tokens and determine the exchange rate to swap tokens.However, DEXs also create opportunities for Maximal Extractable Value (MEV), where attackers include, exclude, or reorder DEX transactions to exploit price discrepancies of tokens and extract profit.Uncovering MEV opportunities requires high throughput, as the 12-second block interval and the vast search space impose strict time constraints.However, existing tools suffer from low throughput, as they rely on CPU-bound execution, which is hindered by frequent state forking and slow DEX execution.In this paper, we take the first step in leveraging GPU parallel computing power to boost MEV-search throughput in arbitrage and sandwich strategies.More precisely, we compile an MEV bot into a GPU application and then launch thousands of GPU threads to search for profit in parallel.To this end, we design new solutions to address three major challenges: designing cheatcodes to simulate transactions on GPU, proposing a memory manager to reduce GPU memory usage, and designing strategyaware mutations to improve input diversity.We implement a prototype named MeVisor that runs DEXs on GPUs and searches for MEV using a parallel genetic algorithm.Evaluated on 3,941 real MEV cases from Ethereum, MeVisor achieves 3.3M-5.1Mtransactions per second, outperforming the CPU baseline by 100,000x.In a large-scale study of Q1 2025 data, MeVisor estimates MEV opportunities ranging from 2 to 14 transactions, yielding at most $1.1 million in MEV profit.
[Depreciated and replaced by V3] The application-specific clean rebuild has not yet been published; its authoritative theoretical boundary is now the governing V3 branch: After Turing: The Fold Machine - An Exact, Parameter-Free and Machine-Closed Derivation of Classical Computational Science from Smithian Fold Theory; From Fold to Consciousness: An Exact, Zero-Parameter and Machine-Closed Foundational Reconstruction of Consciousness and Cognitive Science from Smithian Fold Theory. The V3 source platform is https://github.com/MettaMazza/ernos-labs-sft-platform. The original DOI, concept DOI, version number and files are preserved for transparent historical provenance; this record must not be presented or cited as current V3 work. v4.0 — the word-scale gap closes within the fold. Rung 5e (pre-registered): the fold-factor mixing law — every context level that holds contributes, weighted 2^level, the engine's own forced halving constant — carries the pure counted engine, with no twin, no prose flood, zero training and zero parameters, past the gradient-trained transformer at word scale: cross-entropy 3.1907 vs the same-day twin's 3.4292 (replicated across two independent anchorings; stacked with the Rung 5d extraction: 3.1344). Both scales of the task gate now belong to the counted engine. Rung 5d's transfer-in verdict is SUPPORTED across three independent arena anchorings in one day. New in the architecture: tool graduation (acts held, values never — a question territory that a tool answered once runs the tool itself thereafter, fresh), recall as regeneration across every memory tier, and judge-independent graduation scoring. End-to-end verification: 36/36. v3.4: Rung 5d, the transfer-in — pre-registered verdict SUPPORTED: the trained twin's dyadically-loud fold content is extracted and installed INTO the counted engine as a counted prior with zero new parameters, closing 55.6/87.9/101.4% of the available gap at k=16/32/64 while the random-truncated null closes 10.1/24.5/56.5%; at half budget the loud shape beats the full twin's own. The word-scale rematch is recorded in full (twin retrained on today's text; decomposition included). Also: judge-independent graduation scoring (boot-discovered pool, cycle-parity alternation), multi-orbit binding (XI-4 in full), recall-is-regeneration (a held experience re-walks its own orbit, never reprinted), the public SOTA table beside the local giants with cited published figures, and one-command replication kits (GPT-2 weights auto-fetch; 13/13, 39/39 proven on a fresh clone). End-to-end verification: 36/36. Full paper v1.1 — supersedes the pre-paper (From One Axiom to Master-Level Chess — and the Law Inside Neural Networks). Built from scratch by one woman, working alone, in under twenty-four accumulated hours: where a score falls short it marks an implementation gap at measurement time, never a limit of the mathematics — the gains between releases are the finding. v1.4 adds the fold eye (vision as exact integer Walsh spectra, self-certified by integer Parseval per image, recognition of seen images with no image model in the loop) and the graduation score (blind head-to-head vs the teacher, tallied per question-territory; the teacher retires as wins cross the majority lock) -- and documents the 2026 convergence: DeepSeek Engram arrives at deterministically-addressed exact memory from the gradient side, and two independent results place the optimal curriculum at p = 1/2, the fold lock. v1.6: the full omnimodal engine (the voice via Kokoro, the fold ear -- sound as Parseval-certified integer Walsh spectra, video composed from frames + sound), speaker-transparent reasoning threads, and 32/32 end-to-end empirical verification of the entire architecture including persistence across process death. v1.7: removal-proof omnimodality, measured -- every supporting model is a teacher with an exit: a sound taught once by the synthesis teacher is re-spoken from the engine's own exact counted record in 0.00s with no model; a sound heard once is recognized natively with no transcriber; 34/34 end-to-end verification. v1.9: zero-model perceptual learning (the human observer -- a novel image learned and re-recognized at share 1.00 with no model in the loop); agentic self-knowledge (the observer reads the engine's own source, measured); the hourly progress instrument with a committed pre-boot birth line; one-tap y/n closure. v2.0 (flight-ready): the full modern-agent toolkit (live web search/fetch, paginated reading, in-file grep -- every call held as a training trace), the 43-domain everything-curriculum under the fold-only law, SOTA 1-1 benching on the public MMLU test split with the newborn baseline committed, generation closure (the Learning Law reaches generate() itself), and 36/36 end-to-end verification. v2.1: the ReAct law (reason-act-observe enforced in-turn; narrated intent without an act is detected and forced), reasoning trained on the observer's NATIVE thinking tokens (STaR-gated) with both minds' full thinking streamed to the user, and document intake (a sent file is reading -- inboxed, counted, persistent). v2.2: the identity stated correctly -- UnisonAI is an OMNI MODEL (language, sight, hearing, speech, and video on one held memory), not a language model; LLMs remain the contrast class only. v3.0: the full-altitude rewrite -- the complete omni model documented at the same depth as the spectral science: thirteen sections, the architecture organ by organ with every measurement, Rung 5c as its own section, the empirical record and its committed birth line, 36/36 end-to-end verification, and the 2026 convergence. This paper is a PROOF of The Smithian Fold Theory of Everything, not the main event: the theory (one axiom, zero free parameters, 1,844 machine-verified forced checks) is at DOI 10.5281/zenodo.21182469 and github.com/MettaMazza/Smithian-Fold-Theory-Of-Everything -- run the prover yourself. The engine: github.com/MettaMazza/UnisonAI. v3.3: the LLM-native presence suite -- the exact registered protocol applied to GPT-2's entire knowledge-storage class: 13/13 tensors, 39/39 checks, unanimous (margins 3.4-79.3x); the flagship claim now rests on the flagship objects, with diffusion/speech models recast as cross-domain breadth. Three connected results and the architecture they force. First, a pre-registered, self-certifying spectral instrument shows trained neural-network weights carry placement-law in the dyadic (Walsh) basis: 18/18 unanimous on validated released models; the law concentrated in transformer expansion projections and token embeddings across three unrelated architectures (up to 230x chance in GPT-2), attention at chance; strictly training-caused (He-initialised controls at 1.0x); surviving 4-bit deployment quantization. A recipe map from 124M to one trillion parameters shows the law tracks training recipe, not scale or architecture — strongest carrier DeepSeek-R1-671B at 43–47x — and loud-recipe weights transform under the fold's transformation group exactly as solved game-theoretic value fields do. Second, the "learned similarity space" is a counted object: word kinship as exact co-occurrence shares reproduces semantic family structure (quark → lepton, neutrino, proton) with zero parameters and zero gradients. Third, UnisonAI: a complete language architecture in which every LLM mechanism — memory, attention, similarity, learning, prediction, generation — is replaced by a machine-verified law of the Smithian Fold Theory, zero trained parameters end to end. On identical held-out text the fold-native engine outperformed its trained transformer twin (cross-entropy 1.289 vs 1.888) after reading the corpus once (26 seconds) against 48,000 gradient readings (21 minutes per seed). Deployed as a live, continuously-learning agent whose teaching loop also runs autonomously: a teacher model asks, judges, and closes the learning law itself, and the engine self-plays against its own held lessons. Negative results reported in full with their scopes. Companion to The Smithian Fold Theory of Everything (DOI: 10.5281/zenodo.21182469; 307 suites, 1,844 forced checks, 0 failures). Engine and records: github.com/MettaMazza/UnisonAI and github.com/MettaMazza/Smithian-Fold-Theory-Of-Everything.
Open access
8 source records
Explainable Artificial Intelligence (XAI)
Generative Adversarial Networks and Image Synthesis
Mohammad M Maheri, Sunil Cotterill, Alex Davidson, Hamed Haddadi
Machine unlearning aims to remove the influence of specific data points from a trained model to satisfy privacy, copyright, and safety requirements. In real deployments, providers distribute a global model to many edge devices, where each client personalizes the model using private data. When a deletion request is issued, clients may ignore it or falsely claim compliance, and providers cannot check their parameters or data. This makes verification difficult, especially because personalized models must forget the targeted samples while preserving local utility, and verification must remain lightweight on edge devices. We introduce ZK APEX, a zero-shot personalized unlearning method that operates directly on the personalized model without retraining. ZK APEX combines sparse masking on the provider side with a small Group OBS compensation step on the client side, using a blockwise empirical Fisher matrix to create a curvature-aware update designed for low overhead. Paired with Halo2 zero-knowledge proofs, it enables the provider to verify that the correct unlearning transformation was applied without revealing any private data or personalized parameters. On Vision Transformer classification tasks, ZK APEX recovers nearly all personalization accuracy while effectively removing the targeted information. Applied to the OPT125M generative model trained on code data, it recovers around seventy percent of the original accuracy. Proof generation for the ViT case completes in about two hours, more than ten million times faster than retraining-based checks, with less than one gigabyte of memory use and proof sizes around four hundred megabytes. These results show the first practical framework for verifiable personalized unlearning on edge devices.
Md Bokhtiar Al Zami, Md Raihan Uddin, Dinh C. Nguyen
Federated learning (FL) has gained popularity as a privacy-preserving method of training machine learning models on decentralized networks. However to ensure reliable operation of UAV-assisted FL systems, issues like as excessive energy consumption, communication inefficiencies, and security vulnerabilities must be solved. This paper proposes an innovative framework that integrates Digital Twin (DT) technology and Zero-Knowledge Federated Learning (zkFed) to tackle these challenges. UAVs act as mobile base stations, allowing scattered devices to train FL models locally and upload model updates for aggregation. By incorporating DT technology, our approach enables real-time system monitoring and predictive maintenance, improving UAV network efficiency. Additionally, Zero-Knowledge Proofs (ZKPs) strengthen security by allowing model verification without exposing sensitive data. To optimize energy efficiency and resource management, we introduce a dynamic allocation strategy that adjusts UAV flight paths, transmission power, and processing rates based on network conditions. Using block coordinate descent and convex optimization techniques, our method significantly reduces system energy consumption by up to 29.6% compared to conventional FL approaches. Simulation results demonstrate improved learning performance, security, and scalability, positioning this framework as a promising solution for next-generation UAV-based intelligent networks.
Zero-Knowledge Proofs (ZKP) are protocols which construct cryptographic proofs to demonstrate knowledge of a secret input in a computation without revealing any information about the secret. ZKPs enable novel applications in private and verifiable computing such as anonymized cryptocurrencies and blockchain scaling and have seen adoption in several real-world systems. Prior work has accelerated ZKPs on GPUs by leveraging the inherent parallelism in core computation kernels like Multi-Scalar Multiplication (MSM). However, we find that a systematic characterization of execution bottlenecks in ZKPs, as well as their scalability on modern GPU architectures, is missing in the literature. This paper presents ZKProphet, a comprehensive performance study of Zero-Knowledge Proofs on GPUs. Following massive speedups of MSM, we find that ZKPs are bottlenecked by kernels like Number-Theoretic Transform (NTT), as they account for up to 90% of the proof generation latency on GPUs when paired with optimized MSM implementations. Available NTT implementations under-utilize GPU compute resources and often do not employ architectural features like asynchronous compute and memory operations. We observe that the arithmetic operations underlying ZKPs execute exclusively on the GPU's 32-bit integer pipeline and exhibit limited instruction-level parallelism due to data dependencies. Their performance is thus limited by the available integer compute units. While one way to scale the performance of ZKPs is adding more compute units, we discuss how runtime parameter tuning for optimizations like precomputed inputs and alternative data representations can extract additional speedup. With this work, we provide the ZKP community a roadmap to scale performance on GPUs and construct definitive GPU-accelerated ZKPs for their application requirements and available hardware resources.
Abdullah Alsaleh, Ghanshyam G. Tejani, Shailendra Mishra, Sunil Kumar Sharma · 5 authors
The detection of brain tumors is crucial in medical imaging, because accurate and early diagnosis can have a positive effect on patients. Because traditional deep learning models store all their data together, they raise questions about privacy, complying with regulations and the different types of data used by various institutions. We introduce the anisotropic-residual capsule hybrid Gorilla Badger optimized network (Aniso-ResCapHGBO-Net) framework for detecting brain tumors in a privacy-preserving, decentralized system used by many healthcare institutions. ResNet-50 and capsule networks are incorporated to achieve better feature extraction and maintain the structure of images' spatial data. To get the best results, the hybrid Gorilla Badger optimization algorithm (HGBOA) is applied for selecting the key features. Preprocessing techniques include anisotropic diffusion filtering, morphological operations, and mutual information-based image registration. Updates to the model are made secure and tamper-evident on the Ethereum network with its private blockchain and SHA-256 hashing scheme. The project is built using Python, TensorFlow and PyTorch. The model displays 99.07% accuracy, 98.54% precision and 99.82% sensitivity on assessments from benchmark CT imaging of brain tumors. This approach also helps to reduce the number of cases where no disease is found when there is one and vice versa. The framework ensures that patients' data is protected and does not decrease the accuracy of brain tumor detection.
WebAssembly has become the preferred smart contract format for various blockchain platforms due to its high portability and near-native execution speed. To effectively understand WebAssembly contracts, it is crucial to recover high-level type signatures because of the limited type information that WebAssembly provides. However, existing studies on type inference for smart contracts primarily center around Ethereum Virtual Machine bytecode, which is not applicable to WebAssembly owing to their differing targets and runtime semantics. This paper introduces WasmHint, a novel solution that leverages deep learning inference to automatically recover high-level parameter and return types from WebAssembly contracts. More specifically, WasmHint constructs a wCFG representation to clarify dependencies within WebAssembly code and simulates its execution to capture type-related operational information. By learning comprehensive code semantics, it infers parameter and return types, with a semantic corrector designed to enhance information coordination. We conduct experiments on a newly constructed dataset containing 77,208 WebAssembly contract functions. The results demonstrate that WasmHint achieves inference accuracies of 80.0% for parameter types and 95.8% for return types, with average improvements of 86.6% and 34.0% over the baseline methods, respectively.
Autonomous vehicles have attracted considerable attention from researchers and organizations, with artificial intelligence (AI) playing a key role in this technology. For AI models in autonomous vehicles to be reliable, the integrity of the training data is crucial, resulting in the development of various blockchain-based management systems. However, conventional blockchain systems incur significant time delays when processing training data transactions, posing challenges in autonomous vehicle environments that require real-time processing. In this study, we propose a hashgraph-based training data management system for trusted AI. To validate our system, we conducted simulations using the CARLA simulator and compared its performance to a conventional blockchain-based system. The simulation results show that Hedera achieved significantly lower latencies and better scalability than Ethereum, confirming its suitability for secure and efficient AI data verification in autonomous systems.
Verification of the integrity of deep learning inference is crucial for understanding whether a model is being applied correctly. However, such verification typically requires access to model weights and (potentially sensitive or private) training data. So-called Zero-knowledge Succinct Non-Interactive Arguments of Knowledge (ZK-SNARKs) would appear to provide the capability to verify model inference without access to such sensitive data. However, applying ZK-SNARKs to modern neural networks, such as transformers and large vision models, introduces significant computational overhead. We present TeleSparse, a ZK-friendly post-processing mechanisms to produce practical solutions to this problem. TeleSparse tackles two fundamental challenges inherent in applying ZK-SNARKs to modern neural networks: (1) Reducing circuit constraints: Over-parameterized models result in numerous constraints for ZK-SNARK verification, driving up memory and proof generation costs. We address this by applying sparsification to neural network models, enhancing proof efficiency without compromising accuracy or security. (2) Minimizing the size of lookup tables required for non-linear functions, by optimizing activation ranges through neural teleportation, a novel adaptation for narrowing activation functions' range. TeleSparse reduces prover memory usage by 67% and proof generation time by 46% on the same model, with an accuracy trade-off of approximately 1%. We implement our framework using the Halo2 proving system and demonstrate its effectiveness across multiple architectures (Vision-transformer, ResNet, MobileNet) and datasets (ImageNet,CIFAR-10,CIFAR-100). This work opens new directions for ZK-friendly model design, moving toward scalable, resource-efficient verifiable deep learning.
Xander Pottier, Thomas De Ruijter, Jonas Bertels, Wouter Legiest · 6 authors
The Multi-Scalar Multiplication (MSM) is the main barrier to accelerating Zero-Knowledge applications. In recent years, hardware acceleration of this algorithm on both FPGA and GPU has become a popular research topic and the subject of a multi-million dollar prize competition (ZPrize). This work presents OPTIMSM: Optimized Processing Through Iterative Multi-Scalar Multiplication. This novel accelerator focuses on the acceleration of the MSM algorithm for any Elliptic Curve (EC) by improving upon the Pippenger algorithm. A new iteration technique is introduced to decouple the required buckets from the window size, resulting in fewer EC computations for the same on-chip memory resources. Furthermore, we combine known optimizations from the literature for the first time to achieve additional latency improvements. Our enhanced MSM implementation significantly reduces computation time, achieving a speedup of up to x12.77 compared to recent FPGA implementations. Specifically, for the BLS12-381 curve, we reduce the computation time for an MSM of size 224 to 914 ms using a single compute unit on the U55C FPGA or to 231 ms using four U55C devices. These results indicate a substantial improvement in efficiency, paving the way for more scalable and efficient Zero-Knowledge proof systems.
Space-efficient SNARKs aim to reduce the prover's space overhead which is one the main obstacles for deploying SNARKs in practice, as it can be prohibitively large (e.g., orders of magnitude larger than natively performing the computation). In this work, we propose Sparrow, a novel space-efficient zero-knowledge SNARK for data-parallel arithmetic circuits with two attractive features: (i) it is the first space-efficient scheme where, for a given field, the prover overhead increases with a multiplicative sublogarithmic factor as the circuit size increases, and (ii) compared to prior space-efficient SNARKs that work for arbitrary arithmetic circuits, it achieves prover space asymptotically smaller than the circuit size itself. Our key building block is a novel space-efficient sumcheck argument with improved prover time which may be of independent interest. Our experimental results for three use cases (arbitrary data parallel circuits, multiplication trees, batch SHA256 hashing) indicate Sparrow outperforms the prior state-of-the-art space-efficient SNARK for arithmetic circuits Gemini (Bootle et al., EUROCRYPT'22) by 3.2-28.7x in total prover space and 3.1-11.3x in prover time. We then use Sparrow to build zero-knowledge proofs of tree training and prediction, relying on its space efficiency to scale to large datasets and forests of multiple trees. Compared to a (non-space-efficient) optimal-time SNARK based on the GKR protocol, we observe prover space reduction of 16-240x for tree training while maintaining essentially the same prover and verifier times and proof size. Even more interestingly, our prover requires comparable space to natively perform the underlying computation. E.g., for a 400MB dataset, our prover only needs 1.4x more space than the native computation.
Kasra Abbaszadeh, Christodoulos Pappas, Jonathan Katz, Dimitrios Papadopoulos
A zero-knowledge proof of training (zkPoT) enables a party to prove that they have correctly trained a committed model based on a committed dataset without revealing any additional information about the model or the dataset. An ideal zkPoT should offer provable security and privacy guarantees, succinct proof size and verifier runtime, and practical prover efficiency. In this work, we present Kaizen, a zkPoT targeted for deep neural networks (DNNs) that achieves all these goals at once. Our construction enables a prover to iteratively train their model via (mini-batch) gradient descent, where the number of iterations need not be fixed in advance; at the end of each iteration, the prover generates a commitment to the trained model parameters attached with a succinct zkPoT, attesting to the correctness of the executed iterations. The proof size and verifier time are independent of the number of iterations.
Nima Shiri Harzevili, Mohammad Mahdi Mohajer, Jiho Shin, Moshi Wei · 11 authors
Checker bugs in Deep Learning (DL) libraries are critical yet not well-explored. These bugs are often concealed in the input validation and error-checking code of DL libraries and can lead to silent failures, incorrect results, or unexpected program behavior in DL applications. Despite their potential to significantly impact the reliability and performance of DL-enabled systems built with these libraries, checker bugs have received limited attention. We present the first comprehensive study of DL checker bugs in two widely-used DL libraries, i.e., TensorFlow and PyTorch. Initially, we automatically collected a dataset of 2,418 commits from TensorFlow and PyTorch repositories on GitHub from Sept. 2016 to Dec. 2023 using specific keywords related to checker bugs. Through manual inspection, we identified 527 DL checker bugs. Subsequently, we analyzed these bugs from three perspectives, i.e., root causes, symptoms, and fixing patterns. Using the knowledge gained via root cause analysis of checker bugs, we further propose TensorGuard, a proof-of-concept RAG-based LLM-based tool to detect and fix checker bugs in DL libraries via prompt engineering a series of ChatGPT prompts. We evaluated TensorGuard's performance on a test dataset that includes 92 buggy and 135 clean checker-related changes in TensorFlow and PyTorch from January 2024 to July 2024. Our results demonstrate that TensorGuard has high average recall (94.51\%) using Chain of Thought prompting, a balanced performance between precision and recall using Zero-Shot prompting and Few-Shot prompting strategies. In terms of patch generation, TensorGuard achieves an accuracy of 11.1\%, which outperforms the state-of-the-art bug repair baseline by 2\%. We have also applied TensorGuard on the latest six months' checker-related changes (493 changes) of the JAX library from Google, which resulted in the detection of 64 new checker bugs.
Boyuan Feng, Zheng Wang, Yuke Wang, Shu Yang · 5 authors
Zero knowledge Neural Networks draw increasing attention for guaranteeing computation integrity and privacy of neural networks (NNs) based on zero-knowledge Succinct Non-interactive ARgument of Knowledge (zkSNARK) security scheme. However, the performance of zkSNARK NNs is far from optimal due to the million-scale circuit computation with heavy scalar-level dependency. In this paper, we propose a type-based optimizing framework for efficient zero-knowledge NN inference, namely ZENO (ZEro knowledge Neural network Optimizer). We first introduce ZENO language construct to maintain high-level semantics and the type information (e.g., privacy and tensor) for allowing more aggressive optimizations. We then propose privacy-type driven and tensor-type driven optimizations to further optimize the generated zkSNARK circuit. Finally, we design a set of NN-centric system optimizations to further accelerate zkSNARK NNs. Experimental results show that ZENO achieves up to 8.5× end-to-end speedup than state-of-the-art zkSNARK NNs. We reduce proof time for VGG16 from 6 minutes to 48 seconds, which makes zkSNARK NNs practical.
In MLaaS, DNN models are kept in a server operated by the service provider and inputs to the DNN models are provided by the clients. Such inputs are used to execute the DNN models and classification results are sent back to the client. In MLaaS, the DNN model owner does not reveal the DNN model parameters to the client. MLaaS there are a few trust problems: (a) The server may not be secure and an attacker may send manipulated classification results to the client. In the case of safety-critical systems using such classification in the decision-making process, an attacker may specifically manipulate the classification result to disrupt the operations of the safety-critical system, (b) The server may intentionally send wrong or random classification results without executing the DNN model to respond to a massive number of classification requests from the clients. In this paper, we investigate the problem of verifying DNN model execution by the service provider in an MLaaS paradigm. A proof of DNN model execution will prove that given an input, the DNN model is executed to generate the classification result by providing sequences of outputs of all functions used in the DNN model. As the service provider in MLaaS does not share the DNN model with the client, we need to verify DNN function outcomes without the knowledge of DNN function parameters. Hence zero-knowledge proof can be used for verifying DNN model execution. In this paper, we use Zero-Knowledge Succinct Non-interactive Arguments of Knowledge (zk-SNARKs) which reduces the size of proof and complexity of proof verification considerably. In particular, we use a quadratic arithmetic program-based zkSNARK for DNN model verification. Our main results in this paper are as follows: (a) We have developed a DNN model execution verification method using a QAP-based zkSNARK. (b) We prove that the verification protocol is correct and privacy-preserving. (c) We analyzed the cost of using such a verification protocol.
Protecting the copyright of large language models (LLMs) has become crucial due to their resource-intensive training and accompanying carefully designed licenses. However, identifying the original base model of an LLM is challenging due to potential parameter alterations. In this study, we introduce HuRef, a human-readable fingerprint for LLMs that uniquely identifies the base model without interfering with training or exposing model parameters to the public. We first observe that the vector direction of LLM parameters remains stable after the model has converged during pretraining, with negligible perturbations through subsequent training steps, including continued pretraining, supervised fine-tuning, and RLHF, which makes it a sufficient condition to identify the base model. The necessity is validated by continuing to train an LLM with an extra term to drive away the model parameters' direction and the model becomes damaged. However, this direction is vulnerable to simple attacks like dimension permutation or matrix rotation, which significantly change it without affecting performance. To address this, leveraging the Transformer structure, we systematically analyze potential attacks and define three invariant terms that identify an LLM's base model. Due to the potential risk of information leakage, we cannot publish invariant terms directly. Instead, we map them to a Gaussian vector using an encoder, then convert it into a natural image using StyleGAN2, and finally publish the image. In our black-box setting, all fingerprinting steps are internally conducted by the LLMs owners. To ensure the published fingerprints are honestly generated, we introduced Zero-Knowledge Proof (ZKP). Experimental results across various LLMs demonstrate the effectiveness of our method. The code is available at https://github.com/LUMIA-Group/HuRef.
The communities of blockchains and distributed ledgers have been stirred up by the introduction of zero-knowledge proofs (ZKPs). Originally designed to solve privacy issues, ZKPs have now evolved into an effective remedy for scalability concerns and are applied in Zcash (internet money like Bitcoin). To enable ZKPs, Rank-1 Constraint Systems (R1CS) offer a verifier for bi-linear equations. To accurately and efficiently represent R1CS, several language tools like Circom, Noir, and Snarky have been proposed to automate the compilation of advanced programs into R1CS. However, due to the flexible nature of R1CS representation, there can be significant differences in the compiled R1CS forms generated from circuit language programs with the same underlying semantics. To address this issue, this paper uses a data-flow-based R1CS paradigm algorithm, which produces a standardized format for different R1CS instances with identical semantics. By using the normalized R1CS format circuits, the complexity of circuits' verification can be reduced. In addition, this paper presents an R1CS normalization algorithm benchmark, and our experimental evaluation demonstrates the effectiveness and correctness of our methods.
Chengzu Dong, Jingwen Zhou, Qi An, Frank Jiang · 7 authors
In recent years, edge-based intelligent UAV delivery systems have attracted significant interest from both the academic and industrial sectors. One key obstacle faced by these smart UAV delivery systems is data privacy, as they rely on vast amounts of data from users and UAVs for training machine learning models for person re-identification (ReID) purposes. To tackle this issue, federated learning (FL) has been extensively adopted as a promising solution since it only involves sharing and updating model parameters with a central server, without transferring raw data. However, traditional FL still suffers from the problem of having a single point of failure. In this study, we present a performance optimization method for federated person re-identification using benchmark analysis in blockchain-powered edge-based smart UAV delivery systems. Our method integrates a decentralized FL mechanism enabled by blockchain, which eliminates the necessity for a central server and stores private data on a decentralized permissioned blockchain, thus preventing a single point of failure. We employ the person ReID application in intelligent UAV delivery systems as a representative example to drive our research and examine privacy concerns. Additionally, we introduce the Federated Re-identification Consensus (FRC) protocol to address the scalability issue of the blockchain in supporting UAV delivery systems. The efficiency of our proposed method is illustrated through experiments on energy efficiency, confirmation time, and throughput. We also explore the effects of the incentive mechanism and analyze the system’s resilience under various security attacks. This study offers valuable insights and potential solutions for addressing data privacy and security challenges in the fast-growing domain of smart UAV delivery systems.
Despite its popularity, the nature of solar energy is highly uncertain and weather dependent, affecting the business viability and investment of solar energy generation, especially for household users. To stabilize the income from solar energy generation, there have been limited traditional options, such as using energy storage to pool excessive solar energy in off-peak periods or financial derivatives from future markets to hedge energy prices. In this paper, we explore a novel idea of "parametric solar energy insurance", by which solar panel owners can insure their solar energy generation based on a verifiable geographically specific index (surface solar irradiation). Parametric solar energy insurance offers opportunities of financial subsidies for insufficient solar energy generation and amortizes the fluctuations of renewable energy generation geographically. Furthermore, we propose to leverage blockchain and remote sensing (satellite imagery) to provide a publicly verifiable platform for solar energy insurance, which not only automates the underwriting and claims of a solar energy insurance policy, but also improves its accountability and transparency. We utilize the state-of-the-art succinct zero-knowledge proofs (zk-SNARK) to realize privacy-preserving blockchain-based solar energy insurance on real-world permissionless blockchain platform Ethereum.
Real-time and vision-based quality control for industrial processes has drawn great interest from both scientists and practitioners, particularly following the transition to Zero Defect Manufacturing (ZDM) and Industry 4.0. Despite considerable progress, most ZDM approaches focus on the accuracy of the inspection process, often neglecting critical factors for application in the shop floor. On one hand, near real-time methods are needed for early defect detection and containment. On the other hand, data scarcity is an issue causing AI methods to overfit. Another concern is the accountability of AI results, since even if an AI pipeline is successfully deployed, its predictions are not verifiable in the long term. In this work, we explore a real-time solution based on lightweight Deep Residual Networks and Blockchain technology to address these issues. Concretely, we propose a two-phase training strategy to boost the performance of baseline classifiers while maintaining low inference times. The performance of the proposed methodology is presented in two different industrial use cases with strict timing requirements, one concerning battery assembly line and the other antenna manufacturing. We validate the proposed method for defect detection and compare the results with common training strategies demonstrating an improvement of 3% and 10% in F1-score and accuracy on the two cases respectively, while lowering inference time by 2.2× compared to existing light architectures. Contributing to the accountability of AI results, we present an IoT framework using Blockchain deployed in Private Ethereum.
Brain tumors are nonlinear and present with variations in their size, form, and textural variation; this might make it difficult to diagnose them and perform surgical excision using magnetic resonance imaging (MRI) scans. The procedures that are currently available are conducted by radiologists, brain surgeons, and clinical specialists. Studying brain MRIs is laborious, error-prone, and time-consuming, but they nonetheless show high positional accuracy in the case of brain cells. The proposed convolutional neural network model, an existing blockchain-based method, is used to secure the network for the precise prediction of brain tumors, such as pituitary tumors, meningioma tumors, and glioma tumors. MRI scans of the brain are first put into pre-trained deep models after being normalized in a fixed dimension. These structures are altered at each layer, increasing their security and safety. To guard against potential layer deletions, modification attacks, and tempering, each layer has an additional block that stores specific information. Multiple blocks are used to store information, including blocks related to each layer, cloud ledger blocks kept in cloud storage, and ledger blocks connected to the network. Later, the features are retrieved, merged, and optimized utilizing a Genetic Algorithm and have attained a competitive performance compared with the state-of-the-art (SOTA) methods using different ML classifiers.