Latent-space Attacks for Refusal Evasion in Language Models

Safety-aligned language models are trained to refuse harmful requests, yet refusal behavior can be suppressed by steering their internal representations. Existing methods do so by ablating a refusal direction from model activations, aiming to remove refusal from the model's residual stream. Despite their empirical success, these methods lack a principled account of the latent-space transformation they induce and why it suppresses refusal. In this work, we recast refusal suppression as a latent-space evasion attack against linear probes trained to separate refused from answered prompts. Under this view, prior work's difference-in-means direction naturally defines such a probe, and its ablation is exactly a projection onto its decision boundary, i.e., a minimum-confidence evasion attack. This perspective not only explains the empirical success of prior work but also admits a key limitation: evasion stops at the decision boundary, motivating the need to push representations further into the compliant region, i.e., where the model answers. We leverage this by proposing a Controlled Latent-space Evasion attack that projects representations past the boundary with an optimized confidence. We achieve state-of-the-art attack success rate across 15 instruction-tuned, multimodal, and reasoning models, outperforming existing refusal-ablation baselines and specialized jailbreak attacks.

Label-efficient Training Updates for Malware Detection over Time

Machine Learning (ML)-based detectors are becoming essential to counter the proliferation of malware. However, common ML algorithms are not designed to cope with the dynamic nature of real-world settings, where both legitimate and malicious software evolve. This distribution drift causes models trained under static assumptions to degrade over time unless they are continuously updated. Regularly retraining these models, however, is expensive, since labeling new acquired data requires costly manual analysis by security experts. To reduce labeling costs and address distribution drift in malware detection, prior work explored active learning (AL) and semi-supervised learning (SSL) techniques. Yet, existing studies (i) are tightly coupled to specific detector architectures and restricted to a specific malware domain, resulting in non-uniform comparisons; and (ii) lack a consistent methodology for analyzing the distribution drift, despite the critical sensitivity of the malware domain to temporal changes. In this work, we bridge this gap by proposing a model-agnostic framework that evaluates an extensive set of AL and SSL techniques, isolated and combined, for Android and Windows malware detection. We show that these techniques, when combined, can reduce manual annotation costs by up to 90% across both domains while achieving comparable detection performance to full-labeling retraining. We also introduce a methodology for feature-level drift analysis that measures feature stability over time, showing its correlation with the detector performance. Overall, our study provides a detailed understanding of how AL and SSL behave under distribution drift and how they can be successfully combined, offering practical insights for the design of effective detectors over time.

SAGE-5GC: Security-Aware Guidelines for Evaluating Anomaly Detection in the 5G Core Network

Machine learning-based anomaly detection systems are increasingly being adopted in 5G Core networks to monitor complex, high-volume traffic. However, most existing approaches are evaluated under strong assumptions that rarely hold in operational environments, notably the availability of independent and identically distributed (IID) data and the absence of adaptive attackers.In this work, we study the problem of detecting 5G attacks \textit{in the wild}, focusing on realistic deployment settings. We propose a set of Security-Aware Guidelines for Evaluating anomaly detectors in 5G Core Network (SAGE-5GC), driven by domain knowledge and consideration of potential adversarial threats. Using a realistic 5G Core dataset, we first train several anomaly detectors and assess their baseline performance against standard 5GC control-plane cyberattacks targeting PFCP-based network services.We then extend the evaluation to adversarial settings, where an attacker tries to manipulate the observable features of the network traffic to evade detection, under the constraint that the intended functionality of the malicious traffic is preserved. Starting from a selected set of controllable features, we analyze model sensitivity and adversarial robustness through randomized perturbations. Finally, we introduce a practical optimization strategy based on genetic algorithms that operates exclusively on attacker-controllable features and does not require prior knowledge of the underlying detection model. Our experimental results show that adversarially crafted attacks can substantially degrade detection performance, underscoring the need for robust, security-aware evaluation methodologies for anomaly detection in 5G networks deployed in the wild.

SOM Directions are Better than One: Multi-Directional Refusal Suppression in Language Models

Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model's latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work's difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models' internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.

On the Robustness of Adversarial Training Against Uncertainty Attacks

In learning problems, the noise inherent to the task at hand hinders the possibility to infer without a certain degree of uncertainty. Quantifying this uncertainty, regardless of its wide use, assumes high relevance for security-sensitive applications. Within these scenarios, it becomes fundamental to guarantee good (i.e., trustworthy) uncertainty measures, which downstream modules can securely employ to drive the final decision-making process. However, an attacker may be interested in forcing the system to produce either (i) highly uncertain outputs jeopardizing the system's availability or (ii) low uncertainty estimates, making the system accept uncertain samples that would instead require a careful inspection (e.g., human intervention). Therefore, it becomes fundamental to understand how to obtain robust uncertainty estimates against these kinds of attacks. In this work, we reveal both empirically and theoretically that defending against adversarial examples, i.e., carefully perturbed samples that cause misclassification, additionally guarantees a more secure, trustworthy uncertainty estimate under common attack scenarios without the need for an ad-hoc defense strategy. To support our claims, we evaluate multiple adversarial-robust models from the publicly available benchmark RobustBench on the CIFAR-10 and ImageNet datasets.

Adversarial Pruning: A Survey and Benchmark of Pruning Methods for Adversarial Robustness

Recent work has proposed neural network pruning techniques to reduce the size of a network while preserving robustness against adversarial examples, i.e., well-crafted inputs inducing a misclassification. These methods, which we refer to as adversarial pruning methods, involve complex and articulated designs, making it difficult to analyze the differences and establish a fair and accurate comparison. In this work, we overcome these issues by surveying current adversarial pruning methods and proposing a novel taxonomy to categorize them based on two main dimensions: the pipeline, defining when to prune; and the specifics, defining how to prune. We then highlight the limitations of current empirical analyses and propose a novel, fair evaluation benchmark to address them. We finally conduct an empirical re-evaluation of current adversarial pruning methods and discuss the results, highlighting the shared traits of top-performing adversarial pruning methods, as well as common issues. We welcome contributions in our publicly-available benchmark at https://github.com/pralab/AdversarialPruningBenchmark

HO-FMN: Hyperparameter Optimization for Fast Minimum-Norm Attacks

Gradient-based attacks are a primary tool to evaluate robustness of machine-learning models. However, many attacks tend to provide overly-optimistic evaluations as they use fixed loss functions, optimizers, step-size schedulers, and default hyperparameters. In this work, we tackle these limitations by proposing a parametric variation of the well-known fast minimum-norm attack algorithm, whose loss, optimizer, step-size scheduler, and hyperparameters can be dynamically adjusted. We re-evaluate 12 robust models, showing that our attack finds smaller adversarial perturbations without requiring any additional tuning. This also enables reporting adversarial robustness as a function of the perturbation budget, providing a more complete evaluation than that offered by fixed-budget attacks, while remaining efficient. We release our open-source code at https://github.com/pralab/HO-FMN.

Samples on Thin Ice: Re-Evaluating Adversarial Pruning of Neural Networks

Neural network pruning has shown to be an effective technique for reducing the network size, trading desirable properties like generalization and robustness to adversarial attacks for higher sparsity. Recent work has claimed that adversarial pruning methods can produce sparse networks while also preserving robustness to adversarial examples. In this work, we first re-evaluate three state-of-the-art adversarial pruning methods, showing that their robustness was indeed overestimated. We then compare pruned and dense versions of the same models, discovering that samples on thin ice, i.e., closer to the unpruned model's decision boundary, are typically misclassified after pruning. We conclude by discussing how this intuition may lead to designing more effective adversarial pruning methods in future work.

Improving Fast Minimum-Norm Attacks with Hyperparameter Optimization

Evaluating the adversarial robustness of machine learning models using gradient-based attacks is challenging. In this work, we show that hyperparameter optimization can improve fast minimum-norm attacks by automating the selection of the loss function, the optimizer and the step-size scheduler, along with the corresponding hyperparameters. Our extensive evaluation involving several robust models demonstrates the improved efficacy of fast minimum-norm attacks when hyper-up with hyperparameter optimization. We release our open-source code at https://github.com/pralab/HO-FMN.

Adversarial Attacks Against Uncertainty Quantification

Machine-learning models can be fooled by adversarial examples, i.e., carefully-crafted input perturbations that force models to output wrong predictions. While uncertainty quantification has been recently proposed to detect adversarial inputs, under the assumption that such attacks exhibit a higher prediction uncertainty than pristine data, it has been shown that adaptive attacks specifically aimed at reducing also the uncertainty estimate can easily bypass this defense mechanism. In this work, we focus on a different adversarial scenario in which the attacker is still interested in manipulating the uncertainty estimate, but regardless of the correctness of the prediction; in particular, the goal is to undermine the use of machine-learning models when their outputs are consumed by a downstream module or by a human operator. Following such direction, we: \textit{(i)} design a threat model for attacks targeting uncertainty quantification; \textit{(ii)} devise different attack strategies on conceptually different UQ techniques spanning for both classification and semantic segmentation problems; \textit{(iii)} conduct a first complete and extensive analysis to compare the differences between some of the most employed UQ approaches under attack. Our extensive experimental analysis shows that our attacks are more effective in manipulating uncertainty quantification measures than attacks aimed to also induce misclassifications.