Edoardo Manino

Edoardo Manino, Julia Rozanova, Danilo Carvalho et al.

Metamorphic testing has recently been used to check the safety of neural NLP models. Its main advantage is that it does not rely on a ground truth to generate test cases. However, existing studies are mostly concerned with robustness-like metamorphic relations, limiting the scope of linguistic properties they can test. We propose three new classes of metamorphic relations, which address the properties of systematicity, compositionality and transitivity. Unlike robustness, our relations are defined over multiple source inputs, thus increasing the number of test cases that we can produce by a polynomial factor. With them, we test the internal consistency of state-of-the-art NLP models, and show that they do not always behave according to their expected linguistic properties. Lastly, we introduce a novel graphical notation that efficiently summarises the inner structure of metamorphic relations.

João Batista P. Matos, Iury Bessa, Edoardo Manino et al.

Neural networks are essential components of learning-based software systems. However, their high compute, memory, and power requirements make using them in low resources domains challenging. For this reason, neural networks are often quantized before deployment. Existing quantization techniques tend to degrade the network accuracy. We propose Counter-Example Guided Neural Network Quantization Refinement (CEG4N). This technique combines search-based quantization and equivalence verification: the former minimizes the computational requirements, while the latter guarantees that the network's output does not change after quantization. We evaluate CEG4N~on a diverse set of benchmarks, including large and small networks. Our technique successfully quantizes the networks in our evaluation while producing models with up to 72% better accuracy than state-of-the-art techniques.

Edoardo Manino, Iury Bessa, Lucas Cordeiro

Neural networks are a powerful class of non-linear functions. However, their black-box nature makes it difficult to explain their behaviour and certify their safety. Abstraction techniques address this challenge by transforming the neural network into a simpler, over-approximated function. Unfortunately, existing abstraction techniques are slack, which limits their applicability to small local regions of the input domain. In this paper, we propose Global Interval Neural Network Abstractions with Center-Exact Reconstruction (GINNACER). Our novel abstraction technique produces sound over-approximation bounds over the whole input domain while guaranteeing exact reconstructions for any given local input. Our experiments show that GINNACER is several orders of magnitude tighter than state-of-the-art global abstraction techniques, while being competitive with local ones.

Tong Wu, Edoardo Manino, Fatimah Aljaafari et al.

We describe and evaluate LF-checker, a metaverifier tool based on machine learning. It extracts multiple features of the program under test and predicts the optimal configuration (flags) of a bounded model checker with a decision tree. Our current work is specialised in concurrency verification and employs ESBMC as a back-end verification engine. In the paper, we demonstrate that LF-checker achieves better results than the default configuration of the underlying verification engine.

Philipp Kern, Lorenzo Rovida, Samuel Teuber et al.

Fully homomorphic encryption (FHE) enables private inference by evaluating neural networks on encrypted data. In this way, we can delegate the computation to a third party server without ever revealing the user's data. Currently, the CKKS scheme is the backbone of most efficient FHE implementations, but it only supports addition, multiplication, and array rotation operations, thus requiring all activation functions of the neural network to be approximated by polynomials within a certain interval, imposing strict design tolerances. In this paper, we demonstrate for the first time that this scheme is vulnerable to overflow attacks, i.e., seemingly benign inputs that can exceed such tolerances of the FHE circuit, thereby causing corrupt and unusable outputs. To avoid them, we propose a formal verification technique that computes certified bounds on the ranges of all neurons in the network. By construction, our method eliminates overflows and, in our experiments, removed observed overflows on all benchmarks, reducing failure rates from up to 47% to 0%. Moreover, our overflow-free solution is compatible with most CKKS-based frameworks, as it allows to simply substitute standard polynomials by polynomials with rigorously designed ranges.

Edoardo Manino, Rafael Sá Menezes, Fedor Shmarov et al.

Safety-critical systems with neural network components require strong guarantees. While existing neural network verification techniques have shown great progress towards this goal, they cannot prove the absence of software faults in the network implementation. This paper presents NeuroCodeBench - a verification benchmark for neural network code written in plain C. It contains 32 neural networks with 607 safety properties divided into 6 categories: maths library, activation functions, error-correcting networks, transfer function approximation, probability density estimation and reinforcement learning. Our preliminary evaluation shows that state-of-the-art software verifiers struggle to provide correct verdicts, due to their incomplete support of the standard C mathematical library and the complexity of larger neural networks.

Danilo S. Carvalho, Edoardo Manino, Julia Rozanova et al.

This work proposes a novel methodology for measuring compositional behavior in contemporary language embedding models. Specifically, we focus on adjectival modifier phenomena in adjective-noun phrases. In recent years, distributional language representation models have demonstrated great practical success. At the same time, the need for interpretability has elicited questions on their intrinsic properties and capabilities. Crucially, distributional models are often inconsistent when dealing with compositional phenomena in natural language, which has significant implications for their safety and fairness. Despite this, most current research on compositionality is directed towards improving their performance on similarity tasks only. This work takes a different approach, introducing three novel tests of compositional behavior inspired by Montague semantics. Our experimental results indicate that current neural language models do not behave according to the expected linguistic theories. This indicates that current language models may lack the capability to capture the semantic properties we evaluated on limited context, or that linguistic theories from Montagovian tradition may not match the expected capabilities of distributional models.

Adrians Skapars, Edoardo Manino, Youcheng Sun et al.

Training large language models (LLMs) requires a substantial investment of time and money. To get a good return on investment, the developers spend considerable effort ensuring that the model never produces harmful and offensive outputs. However, bad-faith actors may still try to slander the reputation of an LLM by publicly reporting a forged output. In this paper, we show that defending against such slander attacks requires reconstructing the input of the forged output or proving that it does not exist. To do so, we propose and evaluate a search based approach for targeted adversarial attacks for LLMs. Our experiments show that we are rarely able to reconstruct the exact input of an arbitrary output, thus demonstrating that LLMs are still vulnerable to slander attacks.

Konstantin Kaulen, Tobias Ladner, Stanley Bak et al.

This report summarizes the 6th International Verification of Neural Networks Competition (VNN-COMP 2025), held as a part of the 8th International Symposium on AI Verification (SAIV), that was collocated with the 37th International Conference on Computer-Aided Verification (CAV). VNN-COMP is held annually to facilitate the fair and objective comparison of state-of-the-art neural network verification tools, encourage the standardization of tool interfaces, and bring together the neural network verification community. To this end, standardized formats for networks (ONNX) and specification (VNN-LIB) were defined, tools were evaluated on equal-cost hardware (using an automatic evaluation pipeline based on AWS instances), and tool parameters were chosen by the participants before the final test sets were made public. In the 2025 iteration, 8 teams participated on a diverse set of 16 regular and 9 extended benchmarks. This report summarizes the rules, benchmarks, participating tools, results, and lessons learned from this iteration of this competition.

Xidan Song, Edoardo Manino, Luiz Sena et al.

QNNVerifier is the first open-source tool for verifying implementations of neural networks that takes into account the finite word-length (i.e. quantization) of their operands. The novel support for quantization is achieved by employing state-of-the-art software model checking (SMC) techniques. It translates the implementation of neural networks to a decidable fragment of first-order logic based on satisfiability modulo theories (SMT). The effects of fixed- and floating-point operations are represented through direct implementations given a hardware-determined precision. Furthermore, QNNVerifier allows to specify bespoke safety properties and verify the resulting model with different verification strategies (incremental and k-induction) and SMT solvers. Finally, QNNVerifier is the first tool that combines invariant inference via interval analysis and discretization of non-linear activation functions to speed up the verification of neural networks by orders of magnitude. A video presentation of QNNVerifier is available at https://youtu.be/7jMgOL41zTY

Lucas C. Cordeiro, Matthew L. Daggitt, Julien Girard-Satabin et al.

Neural network verification is a new and rapidly developing field of research. So far, the main priority has been establishing efficient verification algorithms and tools, while proper support from the programming language perspective has been considered secondary or unimportant. Yet, there is mounting evidence that insights from the programming language community may make a difference in the future development of this domain. In this paper, we formulate neural network verification challenges as programming language challenges and suggest possible future solutions.

Yiannis Charalambous, Edoardo Manino, Lucas C. Cordeiro

The next generation of AI systems requires strong safety guarantees. This report looks at the software implementation of neural networks and related memory safety properties, including NULL pointer deference, out-of-bound access, double-free, and memory leaks. Our goal is to detect these vulnerabilities, and automatically repair them with the help of large language models. To this end, we first expand the size of NeuroCodeBench, an existing dataset of neural network code, to about 81k programs via an automated process of program mutation. Then, we verify the memory safety of the mutated neural network implementations with ESBMC, a state-of-the-art software verifier. Whenever ESBMC spots a vulnerability, we invoke a large language model to repair the source code. For the latest task, we compare the performance of various state-of-the-art prompt engineering techniques, and an iterative approach that repeatedly calls the large language model.

Edoardo Manino, Bruno Farias, Rafael Sá Menezes et al.

The behaviour of neural network components must be proven correct before deployment in safety-critical systems. Unfortunately, existing neural network verification techniques cannot certify the absence of faults at the software level. In this paper, we show how to specify and verify that neural networks are safe, by explicitly reasoning about their floating-point implementation. In doing so, we construct NeuroCodeBench 2.0, a benchmark comprising 912 neural network verification examples that cover activation functions, common layers, and full neural networks of up to 170K parameters. Our verification suite is written in plain C and is compatible with the format of the International Competition on Software Verification (SV-COMP). Thanks to it, we can conduct the first rigorous evaluation of eight state-of-the-art software verifiers on neural network code. The results show that existing automated verification tools can correctly solve an average of 11% of our benchmark, while producing around 3% incorrect verdicts. At the same time, a historical analysis reveals that the release of our benchmark has already had a significantly positive impact on the latter.

Adrians Skapars, Edoardo Manino, Youcheng Sun et al.

The task of reconstructing unknown textual inputs to language models is a fundamental auditing primitive that allows us to assess the model's vulnerability to a range of security issues, including stealing hidden system prompts, detecting backdoors, and leaking private data. Existing inversion works assume access to differing levels of information (e.g. requiring input-output examples, the model parameters, intermediate activations or output logits) but oftentimes fail to fully reconstruct the desired input. In this paper, we present the Sparse One-hot Discrete Adam (SODA) algorithm, a search-based inversion method that can accurately reconstruct the input text, given white-box access to the language model and its output. Our experiments demonstrate for the first time that exact language model inversion is possible on both natural language and random inputs. Indeed, SODA achieves respectively 98% and 79% reconstruction rates on inputs with lengths up to 10 tokens. Furthermore, we show that input length and vocabulary size have a far greater impact on the probability of a successful reconstruction than the size of the language model itself, thus allowing us to scale to models from 33M to 3B parameters.

Luiz Sena, Xidan Song, Erickson Alves et al.

Artificial Neural Networks (ANNs) are being deployed for an increasing number of safety-critical applications, including autonomous cars and medical diagnosis. However, concerns about their reliability have been raised due to their black-box nature and apparent fragility to adversarial attacks. These concerns are amplified when ANNs are deployed on restricted system, which limit the precision of mathematical operations and thus introduce additional quantization errors. Here, we develop and evaluate a novel symbolic verification framework using software model checking (SMC) and satisfiability modulo theories (SMT) to check for vulnerabilities in ANNs. More specifically, we propose several ANN-related optimizations for SMC, including invariant inference via interval analysis, slicing, expression simplifications, and discretization of non-linear activation functions. With this verification framework, we can provide formal guarantees on the safe behavior of ANNs implemented both in floating- and fixed-point arithmetic. In this regard, our verification approach was able to verify and produce adversarial examples for $52$ test cases spanning image classification and general machine learning applications. Furthermore, for small- to medium-sized ANN, our approach completes most of its verification runs in minutes. Moreover, in contrast to most state-of-the-art methods, our approach is not restricted to specific choices regarding activation functions and non-quantized representations. Our experiments show that our approach can analyze larger ANN implementations and substantially reduce the verification time compared to state-of-the-art techniques that use SMT solving.

Edoardo Manino, Long Tran-Thanh, Nicholas R. Jennings

A key challenge in crowdsourcing is inferring the ground truth from noisy and unreliable data. To do so, existing approaches rely on collecting redundant information from the crowd, and aggregating it with some probabilistic method. However, oftentimes such methods are computationally inefficient, are restricted to some specific settings, or lack theoretical guarantees. In this paper, we revisit the problem of binary classification from crowdsourced data. Specifically we propose Streaming Bayesian Inference for Crowdsourcing (SBIC), a new algorithm that does not suffer from any of these limitations. First, SBIC has low complexity and can be used in a real-time online setting. Second, SBIC has the same accuracy as the best state-of-the-art algorithms in all settings. Third, SBIC has provable asymptotic guarantees both in the online and offline settings.

Edoardo Manino, Long Tran-Thanh, Nicholas R. Jennings

Crowdsourcing has been successfully employed in the past as an effective and cheap way to execute classification tasks and has therefore attracted the attention of the research community. However, we still lack a theoretical understanding of how to collect the labels from the crowd in an optimal way. In this paper we focus on the problem of worker allocation and compare two active learning policies proposed in the empirical literature with a uniform allocation of the available budget. To this end we make a thorough mathematical analysis of the problem and derive a new bound on the performance of the system. Furthermore we run extensive simulations in a more realistic scenario and show that our theoretical results hold in practice.