Neural Network Verification

Overview

Neural network verification asks whether a model satisfies precise properties, such as resisting bounded input perturbations or keeping a controller within a safe region. Common approaches include interval and linear bound propagation, abstract interpretation, reachability analysis, and solver-based methods such as mixed-integer programming and branch-and-bound. Researchers are extending these techniques from feed-forward classifiers to transformers, recurrent and graph networks, and closed-loop control systems, while also developing benchmarks, proof systems, and tools for trustworthy execution. A central challenge is balancing sound guarantees and tight bounds against the time and memory required, especially for large models and realistic arithmetic.

Latest papers

141 papers11 papers in the last four weeks, down 35% on the four weeks before. 0.1% of all new papers.

Nov 18, 2026

Zonotopic Analysis of Quantization Error in Neural Networks

Quantized neural networks achieve computational efficiency by reducing precision, but bounding their quantization error remains a challenge due to discrete rounding and clamping operations. We present a zonotope-based abstract interpretation framework for quantized ReLU feed-forward network verification, introducing a unified round-clamp abstract transformer that jointly handles both operations. We extend this framework to constrained zonotopes, which add linear constraints to tighten bounds, and derive sound quantization error representations. We establish soundness and affine-strip optimality of the round-clamp abstract transformer, as well as convex-hull optimality of its complete constrained formulation. On the ACAS Xu benchmarks, our constrained zonotope method reduces average error-bound widths by approximately 70-85% and runtimes by 65-80% compared to QEBVerif's symbolic DRA propagation at the two smallest tested perturbation radii (r=3 and r=6). These improvements are not uniform across bitwidths, and symbolic DRA achieves tighter bounds at larger radii. Our plain zonotope method delivers competitive sound bounds nearly 300 times faster on average.
Nov 18, 2026

The 7th International Verification of Neural Networks Competition (VNN-COMP 2026): Summary and Results

This report summarizes the seventh International Verification of Neural Networks Com- petition (VNN-COMP 2026), held as part of the Symposium on AI Verification (SAIV) co-located with the International Conference on Computer-Aided Verification (CAV) at FLoC 2026 in Lisbon, Portugal. The goal of the competition is to provide an objective comparison of the state-of-the-art methods in neural network verification, in terms of scal- ability and speed. Along this line, we used standard formats (ONNX for neural networks and VNNLIB for specifications), standard hardware (all tools are run by the organizers on AWS), and tool parameters provided by the tool authors. This report summarizes the rules, benchmarks, participating tools, results, and lessons learned from this competition.
Nov 17, 2026

BernBRS: Certified Backward Reachability for Deep Bernstein Neural Networks

Ensuring the safety of complex neural network systems requires mathematically rigorous verification. While backward reachability analysis is a critical tool for identifying safe initial states, computing tight, formal bounds remains notoriously difficult to scale. Recent advancements have demonstrated that DeepBern-Nets neural networks utilizing Bernstein polynomial activation functions possess advantageous mathematical properties that enable exceptionally fast and tight forward reachability analysis. Building upon this efficiency, this paper introduces BernBRS, a highly accurate and scalable framework for the backward reachability analysis of DeepBern-Nets. BernBRS simultaneously computes certified inner and outer approximations of the Backward Reachable Set (BRS). Rather than relying on single convex bounding, our method intelligently refines a cover of the input space by analyzing activation patterns, splitting the input domain precisely where an exact decomposition of the relaxation gap predicts the most significant reduction in error. This targeted refinement enables high levels of scalability to highly complex and massive architectures that remain intractable for existing certified methods. We demonstrate the efficacy of our framework on high-dimensional models, including a linearized closed-loop approximation of the Half-Cheetah controller and a distilled Bernstein student of a Transformer block and we empirically study how these model-level guarantees relate to the original Half-Cheetah and TinyBERT classifier, respectively.
Nov 16, 2026

Hyperlyve: Parallelized Enumeration and Exact Lyapunov Stability Certification for Deep ReLU Networks

We introduce Hyperlyve (Hyperplane Partitioned Lyapunov Verification and Enumeration), a framework for proving the asymptotic stability of nonlinear dynamical systems, including those with neural network controllers. Hyperlyve learns and formally certifies continuous piecewise-affine (CPA) neural Lyapunov functions within a counterexample-guided inductive synthesis (CEGIS) loop. Hyperlyve's key innovation is a parallel algorithm that enumerates the hyperplane arrangement formed by the ReLU activations of deep networks and certifies that Lyapunov conditions hold within each polyhedral region. We derive sound and complete regional tests to verify the asymptotic stability of the exact smooth dynamics (rather than a piecewise-linear approximation) within a compact domain of interest. Furthermore, we formally characterize the computational complexity of the CPA neural Lyapunov verification problem as co-NP-complete, filling a key gap in the literature. Empirical results demonstrate that Hyperlyve enumerates ReLU networks over 150-times faster than competing tools, which contributes to Hyperlyve proving the asymptotic stability of systems that are currently intractable for state-of-the-art verifiers.
Nov 16, 2026

Learning and Verifying Stabilizing Controllers with Structured Neural Control Lyapunov Functions

This work investigates the joint learning of neural controllers and control Lyapunov functions (CLFs) for nonlinear discrete-time systems. We exploit the structural properties of Input Convex Neural Networks and Polyak-Łojasiewicz Networks to guarantee positive definiteness of CLF candidates by construction, reducing the conditions requiring formal verification. We develop a mixed-integer programming framework to formally verify the remaining stability conditions and introduce an iterative region of attraction (ROA) expansion procedure to enlarge the certified ROA without retraining. Experiments on an inverted pendulum and a path-tracking system demonstrate the advantages of structured neural CLFs, particularly PLNets, and illustrate how their structural bounds influence the geometry of the certified regions. Our code is available at: https://github.com/JLedesMart/Structured-Neural-Control-Lyapunov-Functions.
Nov 11, 2026

Fit-LRP: Fitting Linear Relaxations to Polytopes for Neural Network Verification over Non-Convex Regions

State-of-the-art neural network verifiers evaluate linear relaxations over axis-aligned bounding boxes. However, verification domains in cyber-physical systems are rarely boxes. Reachability specifications, obstacle-avoidance constraints, and neural Lyapunov certificates typically yield oblique or non-convex polyhedral regions, over which box-based bounds become extremely loose. This paper proposes Fit-LRP (Fitting Linear Relaxations to Polytopes), a framework the implements exact optimization of affine bounds over non-convex domains (represented by unions of polytopes) by solving a sequence of small, low-dimensional linear programs. Our approach seamlessly enhances the final optimization step of LiRPA and α, ⁣βα,\! β-CROWN branch-and-bound (BaB) algorithms. In BaB verification, approximating oblique constraints with axis-aligned splits causes severe combinatorial blowup and risks non-termination. By optimizing the final affine bound of a nonlinear function over the exact polyhedral state-space constraints our method eliminates box-induced over-approximations, enabling earlier search-tree pruning. On ARCH-COMP and neural Lyapunov benchmarks, Fit-LRP reduces branching on the hardest region by 47×47\times, cuts worst-case verification time by 15×15\times, and with region-wise partitioning, verifies on a single CPU 6.1×6.1\times faster than α, ⁣βα,\! β-CROWN on a GPU.
Nov 11, 2026

Conv-Reluplex: A Verification Framework For Convolution Neural Networks

In recent years, machine learning has demonstrated impressive performance in many real-world tasks, especially in computer vision and natural language processing. However, to apply them in safety-critical systems one needs formal guarantees on the neural network outputs. The Reluplex tool is proposed to verify the safety of deep neural networks (DNNs), and in case the DNN fails to give a correct output, can generate adversarial examples. Since the tool can only handle DNNs, it is necessary to extend the tool to process image data. Therefore, in this paper, we propose the Conv-Reluplex framework, which is designed to verify the convolutional layer and pooling layer in convolutional neural networks (CNNs), and generate adversarial examples when classification is misguided. We conduct several experiments on MNIST to evaluate our approaches. The results show that the original CNN is improved using the adversarial examples generated by our tool, and the precision of classification can be increased significantly.
Nov 9, 2026

FocusBranch: Combinatorial Branch-and-Bound for ℓ0\ell_0 Neural Network Robustness Verification

Verifying the local robustness of neural networks under ℓ0\ell_0 perturbations is inherently combinatorial, as it must reason over all sparse subsets of εε perturbed pixels. This paper presents FocusBranch, a combinatorial branch-and-bound framework for ℓ0\ell_0 robustness verification. The key insight is to make the ℓ0\ell_0-ball decomposable and branchable using a node abstraction, which consists of a focus pixel set S\mathcal{S} and a level ll, permitting at most ε−lε-l pixels outside S\mathcal{S}. This abstraction unifies two branching mechanisms: concrete nodes branch through upper shadows to fix additional perturbed locations, while super nodes branch through covering designs to cover many perturbation cases with fewer child nodes. FocusBranch proves node safety by concretizing linear relaxations over node-constrained ℓ0\ell_0-balls and recursively branches on unknown nodes. Rather than relying on a fixed verification plan, it samples nodes across levels to estimate safe rates and costs, then selects a starting level and branching strategy predicted to minimize verification time. Experiments on fully connected and convolutional classifiers for MNIST, Fashion-MNIST, and CIFAR-10 show that FocusBranch verifies ℓ0\ell_0 robustness effectively across architectures and outperforms state-of-the-art verifiers. The code is available at: https://github.com/simpl-group/FocusBranch.
Nov 8, 2026

Scalable Neural Safety Certification via Monotonicity

Learning-based safety certification offers a promising route for verifying complex dynamical systems when explicit models are unavailable, but existing data-driven methods often scale poorly: they treat the system as a black box, rely on dense discretizations or Lipschitz bounds, and consequently suffer from sample complexity that grows exponentially with dimension. This paper shows that a common structural property of dynamical systems--monotonicity--can be used to break this curse of dimensionality. We develop a data-driven framework for robust safety verification and safe controller synthesis for unknown monotone systems using only simulator queries. Our main theoretical result proves that, for monotone systems with upper-closed unsafe sets, safety is completely characterized by the existence of monotone inductive barrier certificates. This structure reduces certificate verification over continuous state spaces to localized boundary checks over finitely many cells. To learn such certificates, we introduce Max-Monotone Networks, a monotone neural architecture with universal approximation guarantees for monotone functions. We then propose a training and refinement algorithm that, upon successful termination, returns formally valid neural barrier certificates and, under mild partition assumptions, requires only O(n)O(n) simulator queries. Across high-dimensional benchmarks, including oscillator networks with up to 13,65913{,}659 states and traffic networks with 1,0001{,}000 states, our method certifies safety in regimes where prior approaches either do not converge or exceed memory limits. These results demonstrate that exploiting order structure can make neural safety certification both scalable and formally sound.
Nov 3, 2026

Robustness Verification against Patch Attacks via Group-Lasso Semidefinite Relaxation

Robustness verification is essential for ensuring the safety and security of neural networks, but current methods focus heavily on norm-ball attacks, which do not adequately model typical real-world threats that present physical perturbations, known as patch attacks. A naive approach requires computationally prohibitive enumeration of all possible patch positions. To address the challenge, we propose a novel neural network verification method specialized to the patch-attack threat. Instead of enumerating patch locations, we first model the patch attack with group-wise sparsity and formulate its verification methods as a mixed-integer semidefinite program (MISDP) by applying SDP relaxation to the neural network. We then propose three relaxations, including Group-Lasso relaxation, that remove the integer variables while preserving the sparse, group-wise structure of the original attack, enabling substantially efficient verification. Our empirical analysis demonstrates that the proposed relaxations require less than 3% of the computation time of the original MISDP, while the Group-Lasso relaxation incurs only a 0.34% average increase in the objective value over the original formulation. Furthermore, Group-Lasso achieves a verified rate of 91%, compared with under 70% for prior methods.
Nov 3, 2026

Sketchpad: certified exact arithmetic with a small vision-language model

Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter executes arithmetic procedures on a rendered sheet. Each forward pass reads a three-cell view and emits one step: a digit, a carry or borrow, and the next operation. A fixed runtime moves the view and writes no digit of a result. We prove in Lean that the model is then exact on all inputs of any length if and only if it is exact on the finite set of views the runtime presents, that no finite test suffices when steps also read earlier steps, and that the guarantee tolerates bounded numerical perturbation and extends to compositions of procedures. All 2,446 views pass on the deployment device, a laptop GPU running the model in a web browser. The certified model performs signed addition and subtraction, long multiplication and division, square roots and greatest common divisors at arbitrary length. It solved 200 of 200 pre-registered problems in the browser, and 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, of which GPT-6.1 Sol, the strongest of ten other models, solved 217. A symbolic runtime built on these procedures solves, expands and factors polynomials with every number computed by the model, and a multiplicative gate on the adapter recovers the base model exactly when the sheet is not in use.
Nov 2, 2026

Iterated Narrow Networks: Turing Completeness at Width 4 on Convex Input Spaces and the Complexity of Verification

We study the iterates of a fixed ReLU network NN of width 4, that is, the discrete-time system xk+1=N(xk)x_{k+1} = N(x_k) on [0,1]4[0,1]^4. Equivalently, xkx_k is the output of a weight-tied network made of kk copies of NN, or of a recurrent network with no external input, started at x0x_0. We prove that every deterministic Turing machine MM admits such a network NMN_M. The network can be constructed in polynomial time from MM. Each iteration has depth poly(∣M∣)\mathrm{poly}(|M|). The iterates of NMN_M simulate MM exactly on every encoded tape. The number of iterations is proportional to the running time of MM. Moreover, the fourth coordinate irreversibly flags every point of the convex cube [0,1]4[0,1]^4 whose simulation reads an invalid digit. This flag is the certificate that makes hardness reductions sound on convex input regions, the regime of practical verifiers. Convex regions were previously inaccessible because of a continuity obstruction inherent to ReLU networks. The correspondence has three consequences. First, reachability for iterated width-4 ReLU networks spans the full complexity hierarchy along the single axis of the time bound: it is NP-complete for a unary bound, NEXP-complete for a binary bound and RE-complete (undecidable) with no bound. Second, two extra neurons of width suffice to extend the hardness to every common smooth activation: at width 6, the robust verification problem (a promise problem with δ/2δδ/2δ gap) is NP-complete for tanh⁡\tanh, sigmoid, GELU, Swish, ELU, and any polynomial-time-computable continuous non-polynomial activation with a polynomial-time-computable Lipschitz constant. Third, the whole spectrum transfers to discretized Neural ODEs in dimension 5.
Oct 25, 2026

Sparse Sum-of-Squares Layers for Certified Learning

A central limitation of modern machine learning systems is the lack of formal certificates guaranteeing their behavior beyond empirical, sample-based evaluation. Sum-of-squares (SOS) methods provide such guarantees by certifying polynomial inequalities over continuous domains via semidefinite programming (SDP). While prior work has introduced SDP layers as differentiable components in neural networks, repeated learning calls can lose the computational benefit of the forward program's sparse representation. We propose Sparse SOS Layers, differentiable value layers whose backward passes reuse the local coordinates of the forward solve. Under a fixed template and learned objective coefficients, each layer avoids building dense certificates and reuses its sparse structure across training steps. Compared with existing methods, our approach achieves up to 2.7 times speedup in total warm-call time, with nearly 100 times lower memory usage (1.7 vs. 160 GiB) for relatively large problem sizes for the tested dense parameterization. Experiments cover sparse polynomial certification, pendulum control, combinatorial relaxations, robotics projection, oscillator chains, and pose graph synchronization. A design study achieves zero certified violations across unseen problem instances and sizes, while requiring roughly 7 times fewer optimization solves than estimating gradients by perturbing the objective coefficients and re-solving the problem (finite differences). These results show that sparse certificate values can be effectively integrated into learning pipelines when the backward pass respects the verifier's structure.
Oct 22, 2026

Convex Equilibrium Potentials for Monge--Ampère: Curvature Bounds and A Posteriori Evaluation

Equilibrium neural networks define their hidden state as the fixed point of a contraction, so their effective depth is set at evaluation time rather than by the architecture. Using them for fully nonlinear PDEs raises two issues that explicit networks avoid: second derivatives may grow with the effective depth, and the potential is only ever evaluated approximately. We study nonnegative softplus equilibrium networks with contraction factor q<1q<1, a convex class suited to the Monge--Ampère equation. A positivity cancellation in the resolvent shows that the Hessian grows at most like (1−q)−2(1-q)^{-2}, rather than the (1−q)−3(1-q)^{-3} of a direct estimate, and an explicit one-neuron family shows that this exponent is sharp; third derivatives satisfy a sharp (1−q)−3(1-q)^{-3} bound. Residual-based a posteriori estimates then control the state, its input derivatives and, in two dimensions, the Hessian determinant at an approximate equilibrium. Combined with interval arithmetic on complete covers of the domain, they yield rigorous uniform bounds on the Monge--Ampère residual and, through a comparison argument, on the solution error of a fixed trained network. On two manufactured Dirichlet problems on the unit disk, all six trained models are certified with residual below 0.250.25 and boundary error below 0.010.01. Adaptive stopping halves float64 evaluation time at best, whereas certified adaptive evaluation remains costlier than fixed depth.
Oct 22, 2026

Training Mixed-Monotone Neural Networks for Verification Without Set Propagation

Most formal verification approaches for neural networks conservatively enclose the output set of the neural network. However, computing a tight output enclosure is in general computationally hard. We sidestep this challenge with a novel neural network architecture that realizes mixed-monotone functions as the difference of monotone subnetworks. By utilizing the theory on mixed-monotone systems, this reduces the computation of a tight output enclosure from complicated set propagation to two simple forward passes of opposite corners of an input interval. Moreover, because the size of the resulting enclosure is a closed-form function of the weights of the neural network, we can directly minimize it during training. We prove a lower bound on the size of any such enclosure, given by the total variation of the represented function generalized to several inputs, which makes it possible to compare the tightness of the enclosure against the optimum. Our evaluation on regression and dynamical-system benchmarks demonstrates that our training reduces the enclosure size to within 0.5% of the lower bound without sacrificing prediction accuracy. Computing our enclosure is orders of magnitude faster than CROWN and αα-CROWN and on deep networks it is also tighter than zonotope propagation and CROWN, while its size exceeds that of αα-CROWN by only 5.1%.
Oct 20, 2026

Enumeration-Guided Training of Neural Hamilton-Jacobi Safety Certificates

Hamilton-Jacobi (HJ) reachability analysis offers formal safety guarantees for control systems, but training accurate neural certificates remains difficult, particularly when unsafe regions are hard to characterize. We propose a verification-guided training framework that addresses this by using probabilistic enumeration of safety constraints to detect forward-invariance violations, i.e., states where the learned certificate incorrectly classifies unsafe behavior. This produces a compact, targeted set of counterexamples that more precisely captures the failure boundary than inputs generated by standard adversarial attack methods. Crucially, using this representation to fine-tune a pretrained HJ value function yields larger certified regions and reduces certification time across all tested benchmarks. We further introduce a shielding mechanism based on the refined certificates to enforce safety constraints at deployment time. We evaluate the proposed safety filter in a realistic, physics-based Isaac Sim environment and through real-world autonomous racing experiments, achieving zero safety constraint violations compared to a nominal controller and a backup policy based on Linear Quadratic Regulation (LQR). These results demonstrate the effectiveness of the proposed shielding mechanism in ensuring safe operation under high-speed, safety-critical conditions.
Oct 16, 2026

Certified neural approximations for parabolic obstacle problems

Residual-trained neural surrogates carry no error guarantee, a limitation for free-boundary problems such as phase change and contact. We derive guaranteed error bounds for neural approximations of parabolic obstacle problems in first-order state--flux--multiplier form. The main result bounds the flux error of the network and the energy error of its feasible projection max⁡{v,g}\max\{v,g\} by the constitutive residual, the initial mismatch, the complementarity defect, and the PDE residual weighted by cF2/αc_F^2/α, without the coercivity and norm-equivalence constants of the first-order formulation. It extends the Apushkinskaya--Repin majorant for the heat operator to general coefficients, free multipliers and a flux bound. A penetration-consistency result shows that neural triples with arbitrarily small bound exist. Outward-rounded cellwise Taylor-model enclosures turn the bound into a finite computation valid on any adaptive partition. The flux and gradient errors of the projected output are certified below 0.270.27 and 0.330.33 of the solution scale for a one-dimensional Stefan problem and, after adaptive refinement, below one half for all three trained networks on a two-dimensional expanding disk. On an expanding ellipse the gradient bound stays above one half; substitution diagnostics trace the floor of the trained networks to the projected state near the free boundary, and the remaining gap to enclosure overhead. Without a closed-form solution, a validated lower bound replaces the solution norm; there the enclosure is nearly exact, but the trained networks have errors above the solution scale and no useful accuracy is certified. The bounds control a state--flux norm, not the multiplier or the free-boundary location.
Oct 12, 2026

Masked Attention Pooling Is Invariant to Padding and Permutation: A Formal Proof in Lean

Attention-based multiple instance learning (MIL) pools a variable-size bag of instance embeddings into one vector. In batched training, bags are padded to a common length, and the pooled output is expected not to depend on how much padding was added or on the order in which instances are stored. This expectation holds only if padded rows are removed from both the numerator and the softmax normalizer. Common implementation idioms meet it only under side conditions. Zeroing padded features alone does not work. Masking logits with −∞-\infty fails on empty bags and on nonfinite padding. Masking with a large negative constant is exact only after floating-point underflow. We define masked attention pooling with a guarded normalizer and prove in Lean 4 with mathlib that it is invariant under any bijective reindexing of rows and under appending any finite number of masked rows with arbitrary real features and logits. We also prove that values at masked positions are irrelevant and that an all-masked bag maps to zero. The theorems hold for arbitrary finite index types and feature dimensions. Using the jax-lean certificate pipeline, we connect these theorems to the Jaxpr traced from the JAX kernel itself. Lean checks translation certificates for two concrete shapes, identifies the generated functions with the generic specification, and transfers both invariances to the imported Jaxpr's real-arithmetic semantics. All audited theorems depend only on Lean's three standard axioms. The results concern ideal real arithmetic. Floating-point execution, XLA, and the JAX tracer and importer remain outside the proof.
Oct 12, 2026

Certified Robustness via Sparse Second-Order Cone Relaxations

Certified robustness is critical for deploying neural networks in safety-sensitive settings, but computing exact certificates typically requires solving nonconvex optimization problems. Practical verifiers use convex relaxations: linear programming (LP) relaxations are scalable but often too conservative, whereas semidefinite programming (SDP) relaxations can be substantially tighter but are computationally expensive. We introduce a hierarchy of second-order cone programming (SOCP) relaxations that provide stronger certificates than LP while avoiding the cost of global positive-semidefinite constraints. Our construction combines convex-hull relaxations of ReLU activations, selective McCormick lifting over certified variable bounds, and second-order cone constraints induced by 2-by-2 principal minors. Adding these constraints progressively yields a monotone hierarchy of increasingly tight bounds with a controllable computation-tightness trade-off. On fully connected and convolutional networks on MNIST and CIFAR-10, SOCP-based verification improves certified robust accuracy over LP-based verification and is competitive with alpha,beta-CROWN, while remaining orders of magnitude faster than SDP-based verification. These results position selective second-order constraints as a practical middle ground between scalable LP-based and tight SDP-based certification. Finally, we introduce class-optimized robust evaluation (CORE), which applies tighter relaxations only to target classes not certified by cheaper ones. On the fully connected MNIST benchmark, CORE variants match SDP-cert's certified robust accuracy while reducing average verification time by up to 86.30%.
Oct 8, 2026

Neural Network Verification for Deep Joint Source-Channel Coding

Deep joint source-channel coding (DeepJSCC) transmits data end-to-end over wireless channels using a neural encoder-decoder, but reconstruction quality can degrade sharply under adversarial perturbations and channel disturbances; no method formally bounds this degradation for DeepJSCC. We present the first bound-propagation framework for verifying DeepJSCC's decoder, bounding worst-case reconstruction error over a given wireless channel's noise region. Current deep neural network (DNN) verifiers do not support three DeepJSCC decoder components: parametric rectified linear activations (PReLU), transposed convolutions, and Rayleigh fading. We extend state-of-the-art techniques for optimization of linear relaxation in DNN verification for PReLU, replace the transposed convolution with its restricted upsample-then-convolution form, and formulate Rayleigh fading as a structural perturbation prepended directly into the decoder, thereby reducing the dimensionality of the verification problem. We also instantiate Lipschitz-regularized global robustness training, denoted GloRo, improving global robustness and enabling tight certification of DeepJSCC models for the first time. On DeepJSCC model for image transmission, this global robustness training procedure combined with structural encoding lowers the median certified bound by up to 41% and certifies about ten times more safe cases (192 against 19) than GloRo with interval encoding at a 10-degree error in channel estimation. Over-the-air validation with an orthogonal frequency-division multiplexing (OFDM) implementation on software-defined radio devices confirm the certificate holds on real hardware, with a worst observed error on radio link at 0.082 against a certified bound of 0.128.
Oct 6, 2026

zkLLMPoT: Efficient Zero Knowledge Proof of Training for Large Language Models

Auditing the claimed outcomes of large language model (LLM) training is challenging when model weights and training data are private, while cryptographically proving the full training process is prohibitively expensive at Transformer scale. We present zkLLMPoT, a zero-knowledge framework that certifies auditor-defined properties of a trained checkpoint through forward evaluation rather than verification of its optimization trajectory. zkLLMPoT includes 2 phases: 1) The trainer fixes the architecture and the model weights are committed. Then the auditor selects challenge sequences, preventing the trainer from modifying the checkpoint in response to the audit data. 2) Then the trainer proves the objective value attained by the committed model on those sequences. This formulation makes the certification cost independent of the number of training iterations, without revealing model weights or requiring access to private training data. We build on sumcheck- and lookup-based arguments to certify Transformer computations, while supporting next-token loss and task-specific audit objectives. Across four model families, operator-level benchmarks yield proving times of 41-59 seconds for 1.1-1.5B-parameter models and 131 seconds at 13B for the covered operators, with verification below half a second at a sequence length of 512.
Oct 5, 2026

Toward AI Trustworthiness: Finding Analytically Proven Forward-Invariant Sets for AI-Controlled Systems

Neural-network (NN) controllers are increasingly used in nonlinear control systems, but their highly nonlinear behavior makes them difficult to explain and verify, raising trustworthiness concerns in safety- and mission-critical applications. A key step toward certifiable trustworthiness is to find a Forward-Invariant Set (FIS): a state-space region such that any trajectory starting inside remains inside. If the FIS excludes unsafe states, safety can be guaranteed for initial states within it. Finding an analytically proven FIS for a given AI-controlled system with a fixed controller is difficult. We propose a framework that uses an Invertible Neural Network (INN) to transform the original state space into a latent space where a regular-shaped FIS is more likely to exist. We train the INN so that a preferred hyper-rectangular candidate becomes invariant in the latent space, then formally verify it. We prove that, whenever verification succeeds, both the latent-space candidate and its inverse-transformed counterpart in the original state space are analytically proven FISs. We evaluate the approach on 45 AI-controlled systems across three representative control testbeds. Our method finds certified FISs for all 45 systems, whereas an adapted state-of-the-art baseline finds none. It is also faster on 40 of the 45 systems, and the centers of the resulting FISs roughly match domain-expert preferences.
Oct 2, 2026

Certified Mechanistic Edits: Behavioral Guarantees for Skill Removal and Preservation

Mechanistic edits (ablations, weight edits, activation steering) are the standard tools for unlearning a harmful capability from a neural network while preserving useful ones. Current approaches validate their effects only by testing, which can never cover an entire continuous region of inputs. Prior work at the interpretability-verification boundary certifies descriptions of a model: what a circuit computes, or whether it faithfully explains the whole. We instead certify the behavioral effect of an edit: that disabling a circuit removes one skill and provably preserves another, for every input in a region; a feature non-interference guarantee in the information-flow-security sense. We demonstrate such certified edits from toy ReLU networks up to a standard softmax + LayerNorm transformer, proving removal and preservation over continuous embedding-space regions and reaching roughly 9x the input-perturbation dimension an exact solver can handle by switching to sound bound propagation. Furthermore, we prove that no finite deterministic black-box test can certify removal, exhibiting an edit that passes exhaustive testing yet provably fails on a survivor pocket that can be made arbitrarily small. Guarantees hold on small, standard-architecture networks and, like any removal claim, presuppose that the target skill admits a decidable specification, a property which real-world harms may not have.
Sep 30, 2026

Security Properties of Neural Networks as Decision Problems

Certifying a deployed neural network raises decision problems that the verification literature has not classified: whether the model carries a backdoor planted in its training data, whether a fault in its stored parameters can drive it into an unsafe state, whether its output leaks a private part of its input. We formalise eight such problems and classify what we can. The organising observation is a logical one. The function computed by a piecewise linear network, together with all its node values, is definable by a quantifier-free formula of real addition of size linear in the network, so a property of the network is a quantifier-alternation sentence, which Sontag's 1985 theorem places in the polynomial hierarchy at the level of its prefix. Membership results are thus corollaries, and the argument makes plain what they need: that the quantified objects are inputs rather than the network's own parameters. Non-interference, monotonicity and counterfactual fairness have exactly the complexity of network equivalence and of interval verification, all co-NP- complete over ReLU. Detection of backdoor triggers from a quantised alphabet is Sigma_2^P-complete, one level above robustness certification, so it does not reduce to polynomially many robustness queries unless the hierarchy collapses. Inversion resistance is co-NP-complete for every l_p metric, p a fixed positive integer. Quantifying over parameters instead of inputs - the fault model of bit-flip attacks, radiation upsets and analog accelerators - makes verification exists-R-complete already for networks of identity nodes, for which every previously studied problem is in P, and it stays so when each parameter is confined to a box of inverse-polynomial width; the corresponding safety question is forall-R-complete for ReLU.
Sep 29, 2026

Stochastic World Models for Verifying Vision-Based Neural Feedback Systems

Verifying a vision-based neural feedback system requires a model of the observations its controller acts upon. Such a model must capture the variation the sensor produces, while remaining tractable for closed-loop analysis. Generative adversarial networks (GANs) have served as perception surrogates, but they are large, reproduce complex scenes poorly, and are hard to verify. We explore stochastic world models as a richer class of perception surrogates. We train a world model with physically grounded latents, built from operations that standard verifiers bound. It reproduces held-out frames more faithfully than GAN surrogates with up to 130 times as many parameters. To verify these surrogates, we develop a procedure that combines falsification, adaptive refinement, symbolic, and backward analyses. On an emergency braking benchmark with a GAN surrogate, our procedure resolves the entire state space, 38% of which the state-of-the-art verifier left unresolved. On the RGB version of the benchmark, where no verification results have previously been reported, our procedure resolves over 80% of the state space with a world model surrogate.
Sep 28, 2026

CLAD: Constrained Abstract Domain for Neural Network Verification

Neural network verification (NNV) formally verifies that a network satisfies a specified property for all inputs within a defined region. Modern NNV tools employ abstract domains to compute a sound over-approximation of the network's behavior from the given input region, thus the tightness of these abstractions essentially determines efficiency. A long line of increasingly precise domains has been developed, but they all describe the valid input region in the same restrictive way, e.g., an Lp-norm ball. A practical input region is rarely a simple Lp ball, but rather a combination Lp ball with additional constraints. Verifying a network over such a region with existing abstraction produces a loose over-approximation, which results in either failing to verify a property or spurious counterexamples. We introduce Constrained Lagrangian Abstract Domain (CLAD), a new abstract domain that computes a sound over-approximation of neural networks over input regions defined by a combination of convex constraints. CLAD propagates these constraints and tightens bounds over the true feasible region. However, bounding a neuron over the intersection of these constraints has no closed-form solution, so CLAD relaxes each constraint into the objective with a Lagrange multiplier and solves the resulting max-min problem with a projected primal-dual method, alternating a projected gradient step on the input with a multiplier update. CLAD supports any convex constraint with a subgradient, e.g., from automatic differentiation. We evaluate CLAD on 1,944 instances across four convolutional networks with motion-blur structured perturbations with halfspace or L2-ball constraints. On standard unconstrained Linf property, CLAD verifies as many instances as GCPCROWN at a similar runtime. On constrained properties, CLAD verifies 60% more instances than GCPCROWN on L2-ball properties, and 22% more in total.
Sep 28, 2026

Verifying Neural Networks with Reinforcement Learning

Formal verification can play a key role in ensuring the reliability of Deep Neural Networks (DNNs) deployed in safety-critical systems. Modern DNN verifiers employ a branch-and-bound framework, which alternates between branching (splitting into smaller subproblems) and bounding (pruning subproblems) to efficiently explore the verification space. However, existing branching heuristics make greedy decisions based on static scoring functions. They do not anticipate long-term efficiency or leverage the growing availability of verification data to improve performance. This work introduces RSB, a reinforcement learning framework that learns to refine baseline branching heuristics. It trains an actor-critic architecture to maximize cumulative future rewards rather than immediate scores. The actor generates attention weights from observations of raw neuron features and learned graph embeddings, which rescale baseline heuristic scores to guide neuron branching. Evaluation on 600 challenging instances demonstrates that RSB consistently outperforms state-of-the-art branching heuristics, solving 11% more instances while reducing branch exploration by 50%.
Sep 28, 2026

ZonoGPT: Towards An Abstract Domain for Verifying Large GPT Models

Transformer-based models are widely used for reasoning, coding, and multimodal agentic tasks. To provide formal assurance of desirable behaviors, such as robustness, safety, and fairness, neural network verification techniques prove required properties and provide auditable guarantees before deployment. However, prior work remains limited to small or restricted Transformers, and maintaining precision across deep models remains challenging. In this work, we introduce ZonoGpt, an abstract domain for verifying large transformers that maintains a space complexity independent of network depth. ZonoGpt uses a structured zonotope and a generator reduction mechanism to efficiently preserve correlations. To maintain precision, it introduces block-specific fused transformations for Attention and LayerNorm that retain feature relations, along with an affine transform for GELU that preserves generator relations. These mechanisms enable ZonoGpt to be the first approach to verify standard architectures, scaling to official HuggingFace models up to GPT-2 Medium (24 blocks, 300M+ parameters) and successfully verifying 1,339 instances across text and vision tasks.
Sep 25, 2026

AC Power Flow Contingency Analysis Using a Single Deep Neural Network

Contingency analysis using the AC power flow (AC-PF) model is a critical tool for accurate grid security assessment, but its computational burden increases with the number of operating scenarios and outage configurations to evaluate. Recent ML-based approaches typically require outage-specific training data, leading to offline training costs that scale with the number of contingencies. This work proposes a framework that reuses a single ML model trained solely on basecase AC-PF data to estimate post-contingency operating states under arbitrary single-line outages. The proposed approach formulates post-contingency state prediction as a fixed-point iteration. If the ML model is a deep neural network (DNN), we derive sufficient conditions that guarantee convergence and develop semidefinite programming (SDP) formulations to certify these conditions for a given DNN. Numerical tests on the IEEE 118-bus system demonstrate that the proposed SDP formulations are tight, that the certified conditions hold for all tested contingencies, and that the resulting method produces accurate post-contingency state estimates within only a few iterations.
Sep 25, 2026

Learning Provable Neural Network Observer for Uncertain Dynamical Systems

In many safety-critical applications, control of uncertain dynamical systems relies on observers that estimate states and external disturbances. Neural network observers can improve estimation accuracy, but certifying their Lyapunov stability via Linear Matrix Inequality (LMI) constraints leads to large-scale semidefinite programs (SDPs) that are difficult to solve for large networks. To overcome this scalability bottleneck, we propose a novel two-stage training framework for provably stable neural network observers. Our approach decouples the optimization into a point-guided Lyapunov pre-training phase, which rapidly achieves high estimation accuracy and local stability over sampled states, followed by an LMI fine-tuning phase that efficiently satisfies a strict global Lyapunov stability certificate. We provide formal theoretical guarantees for local stability radii and probabilistic coverage over a prescribed compact error-state domain under specified regularity and sampling assumptions. Experiments on nonlinear control benchmarks and X-29 aircraft ablations show that our LMI-certified neural network observers train significantly faster than direct LMI-based methods and generalize robustly across diverse systems, achieving improved tracking accuracy over a range of observer baselines. The code is available at https://github.com/Berry-Myon/LearningNeuralNetworkObserver.