Paul Montague

Shuiqiao Yang, Bao Gia Doan, Paul Montague et al. · cambridge

Graph Neural Networks (GNNs) have achieved tremendous success in many graph mining tasks benefitting from the message passing strategy that fuses the local structure and node features for better graph representation learning. Despite the success of GNNs, and similar to other types of deep neural networks, GNNs are found to be vulnerable to unnoticeable perturbations on both graph structure and node features. Many adversarial attacks have been proposed to disclose the fragility of GNNs under different perturbation strategies to create adversarial examples. However, vulnerability of GNNs to successful backdoor attacks was only shown recently. In this paper, we disclose the TRAP attack, a Transferable GRAPh backdoor attack. The core attack principle is to poison the training dataset with perturbation-based triggers that can lead to an effective and transferable backdoor attack. The perturbation trigger for a graph is generated by performing the perturbation actions on the graph structure via a gradient based score matrix from a surrogate model. Compared with prior works, TRAP attack is different in several ways: i) it exploits a surrogate Graph Convolutional Network (GCN) model to generate perturbation triggers for a blackbox based backdoor attack; ii) it generates sample-specific perturbation triggers which do not have a fixed pattern; and iii) the attack transfers, for the first time in the context of GNNs, to different GNN models when trained with the forged poisoned training dataset. Through extensive evaluations on four real-world datasets, we demonstrate the effectiveness of the TRAP attack to build transferable backdoors in four different popular GNNs using four real-world datasets.

Anh Bui, Trung Le, He Zhao et al. · cambridge

Deep learning models, even the-state-of-the-art ones, are highly vulnerable to adversarial examples. Adversarial training is one of the most efficient methods to improve the model's robustness. The key factor for the success of adversarial training is the capability to generate qualified and divergent adversarial examples which satisfy some objectives/goals (e.g., finding adversarial examples that maximize the model losses for simultaneously attacking multiple models). Therefore, multi-objective optimization (MOO) is a natural tool for adversarial example generation to achieve multiple objectives/goals simultaneously. However, we observe that a naive application of MOO tends to maximize all objectives/goals equally, without caring if an objective/goal has been achieved yet. This leads to useless effort to further improve the goal-achieved tasks, while putting less focus on the goal-unachieved tasks. In this paper, we propose \emph{Task Oriented MOO} to address this issue, in the context where we can explicitly define the goal achievement for a task. Our principle is to only maintain the goal-achieved tasks, while letting the optimizer spend more effort on improving the goal-unachieved tasks. We conduct comprehensive experiments for our Task Oriented MOO on various adversarial example generation schemes. The experimental results firmly demonstrate the merit of our proposed approach. Our code is available at \url{https://github.com/tuananhbui89/TAMOO}.

Andrew C. Cullen, Paul Montague, Shijie Liu et al. · cambridge

In response to subtle adversarial examples flipping classifications of neural network models, recent research has promoted certified robustness as a solution. There, invariance of predictions to all norm-bounded attacks is achieved through randomised smoothing of network inputs. Today's state-of-the-art certifications make optimal use of the class output scores at the input instance under test: no better radius of certification (under the $L_2$ norm) is possible given only these score. However, it is an open question as to whether such lower bounds can be improved using local information around the instance under test. In this work, we demonstrate how today's "optimal" certificates can be improved by exploiting both the transitivity of certifications, and the geometry of the input space, giving rise to what we term Geometrically-Informed Certified Robustness. By considering the smallest distance to points on the boundary of a set of certifications this approach improves certifications for more than $80\%$ of Tiny-Imagenet instances, yielding an on average $5 \%$ increase in the associated certification. When incorporating training time processes that enhance the certified radius, our technique shows even more promising results, with a uniform $4$ percentage point increase in the achieved certified radius.

Shijie Liu, Andrew C. Cullen, Paul Montague et al. · cambridge

Poisoning attacks can disproportionately influence model behaviour by making small changes to the training corpus. While defences against specific poisoning attacks do exist, they in general do not provide any guarantees, leaving them potentially countered by novel attacks. In contrast, by examining worst-case behaviours Certified Defences make it possible to provide guarantees of the robustness of a sample against adversarial attacks modifying a finite number of training samples, known as pointwise certification. We achieve this by exploiting both Differential Privacy and the Sampled Gaussian Mechanism to ensure the invariance of prediction for each testing instance against finite numbers of poisoned examples. In doing so, our model provides guarantees of adversarial robustness that are more than twice as large as those provided by prior certifications.

Andrew C. Cullen, Paul Montague, Shijie Liu et al. · cambridge

Certified robustness circumvents the fragility of defences against adversarial attacks, by endowing model predictions with guarantees of class invariance for attacks up to a calculated size. While there is value in these certifications, the techniques through which we assess their performance do not present a proper accounting of their strengths and weaknesses, as their analysis has eschewed consideration of performance over individual samples in favour of aggregated measures. By considering the potential output space of certified models, this work presents two distinct approaches to improve the analysis of certification mechanisms, that allow for both dataset-independent and dataset-dependent measures of certification performance. Embracing such a perspective uncovers new certification approaches, which have the potential to more than double the achievable radius of certification, relative to current state-of-the-art. Empirical evaluation verifies that our new approach can certify $9\%$ more samples at noise scale $σ= 1$, with greater relative improvements observed as the difficulty of the predictive task increases.

Andrew C. Cullen, Shijie Liu, Paul Montague et al. · cambridge

In guaranteeing the absence of adversarial examples in an instance's neighbourhood, certification mechanisms play an important role in demonstrating neural net robustness. In this paper, we ask if these certifications can compromise the very models they help to protect? Our new \emph{Certification Aware Attack} exploits certifications to produce computationally efficient norm-minimising adversarial examples $74 \%$ more often than comparable attacks, while reducing the median perturbation norm by more than $10\%$. While these attacks can be used to assess the tightness of certification bounds, they also highlight that releasing certifications can paradoxically reduce security.

Shigang Liu, Di Cao, Junae Kim et al.

Recently, deep learning has demonstrated promising results in enhancing the accuracy of vulnerability detection and identifying vulnerabilities in software. However, these techniques are still vulnerable to attacks. Adversarial examples can exploit vulnerabilities within deep neural networks, posing a significant threat to system security. This study showcases the susceptibility of deep learning models to adversarial attacks, which can achieve 100% attack success rate (refer to Table 5). The proposed method, EaTVul, encompasses six stages: identification of important samples using support vector machines, identification of important features using the attention mechanism, generation of adversarial data based on these features using ChatGPT, preparation of an adversarial attack pool, selection of seed data using a fuzzy genetic algorithm, and the execution of an evasion attack. Extensive experiments demonstrate the effectiveness of EaTVul, achieving an attack success rate of more than 83% when the snippet size is greater than 2. Furthermore, in most cases with a snippet size of 4, EaTVul achieves a 100% attack success rate. The findings of this research emphasize the necessity of robust defenses against adversarial attacks in software vulnerability detection.

Anh Bui, Long Vuong, Khanh Doan et al.

Diffusion models excel at generating visually striking content from text but can inadvertently produce undesirable or harmful content when trained on unfiltered internet data. A practical solution is to selectively removing target concepts from the model, but this may impact the remaining concepts. Prior approaches have tried to balance this by introducing a loss term to preserve neutral content or a regularization term to minimize changes in the model parameters, yet resolving this trade-off remains challenging. In this work, we propose to identify and preserving concepts most affected by parameter changes, termed as \textit{adversarial concepts}. This approach ensures stable erasure with minimal impact on the other concepts. We demonstrate the effectiveness of our method using the Stable Diffusion model, showing that it outperforms state-of-the-art erasure methods in eliminating unwanted content while maintaining the integrity of other unrelated elements. Our code is available at https://github.com/tuananhbui89/Erasing-Adversarial-Preservation.

Anh Bui, Trang Vu, Long Vuong et al.

Concept erasure has emerged as a promising technique for mitigating the risk of harmful content generation in diffusion models by selectively unlearning undesirable concepts. The common principle of previous works to remove a specific concept is to map it to a fixed generic concept, such as a neutral concept or just an empty text prompt. In this paper, we demonstrate that this fixed-target strategy is suboptimal, as it fails to account for the impact of erasing one concept on the others. To address this limitation, we model the concept space as a graph and empirically analyze the effects of erasing one concept on the remaining concepts. Our analysis uncovers intriguing geometric properties of the concept space, where the influence of erasing a concept is confined to a local region. Building on this insight, we propose the Adaptive Guided Erasure (AGE) method, which \emph{dynamically} selects optimal target concepts tailored to each undesirable concept, minimizing unintended side effects. Experimental results show that AGE significantly outperforms state-of-the-art erasure methods on preserving unrelated concepts while maintaining effective erasure performance. Our code is published at {https://github.com/tuananhbui89/Adaptive-Guided-Erasure}.

Sean Lamont, Christian Walder, Paul Montague et al.

Diverse outputs in text generation are necessary for effective exploration in complex reasoning tasks, such as code generation and mathematical problem solving. Such Pass@$k$ problems benefit from distinct candidates covering the solution space. However, traditional sampling approaches often waste computational resources on repetitive failure modes. While Diffusion Language Models have emerged as a competitive alternative to the prevailing Autoregressive paradigm, they remain susceptible to this redundancy, with independent samples frequently collapsing into similar modes. To address this, we propose a training free, low cost intervention to enhance generative diversity in Diffusion Language Models. Our approach modifies intermediate samples in a batch sequentially, where each sample is repelled from the feature space of previous samples, actively penalising redundancy. Unlike prior methods that require retraining or beam search, our strategy incurs negligible computational overhead, while ensuring that each sample contributes a unique perspective to the batch. We evaluate our method on the HumanEval and GSM8K benchmarks using the LLaDA-8B-Instruct model. Our results demonstrate significantly improved diversity and Pass@$k$ performance across various temperature settings. As a simple modification to the sampling process, our method offers an immediate, low-cost improvement for current and future Diffusion Language Models in tasks that benefit from diverse solution search. We make our code available at https://github.com/sean-lamont/odd.

Sean Lamont, Christian Walder, Amir Dezfouli et al.

A key challenge in automated formal reasoning is the intractable search space, which grows exponentially with the depth of the proof. This branching is caused by the large number of candidate proof tactics which can be applied to a given goal. Nonetheless, many of these tactics are semantically similar or lead to an execution error, wasting valuable resources in both cases. We address the problem of effectively pruning this search, using only synthetic data generated from previous proof attempts. We first demonstrate that it is possible to generate semantically aware tactic representations which capture the effect on the proving environment, likelihood of success, and execution time. We then propose a novel filtering mechanism which leverages these representations to select semantically diverse and high quality tactics, using Determinantal Point Processes. Our approach, 3D- Prover, is designed to be general, and to augment any underlying tactic generator. We demonstrate the effectiveness of 3D-Prover on the miniF2F and LeanDojo benchmarks by augmenting popular open source proving LLMs. We show that our approach leads to an increase in the overall proof rate, as well as a significant improvement in the tactic success rate, execution time and diversity. We make our code available at https://github.com/sean-lamont/3D-Prover.

Anh Bui, Trung Le, He Zhao et al.

Ensemble-based adversarial training is a principled approach to achieve robustness against adversarial attacks. An important technique of this approach is to control the transferability of adversarial examples among ensemble members. We propose in this work a simple yet effective strategy to collaborate among committee models of an ensemble model. This is achieved via the secure and insecure sets defined for each model member on a given sample, hence help us to quantify and regularize the transferability. Consequently, our proposed framework provides the flexibility to reduce the adversarial transferability as well as to promote the diversity of ensemble members, which are two crucial factors for better robustness in our ensemble approach. We conduct extensive and comprehensive experiments to demonstrate that our proposed method outperforms the state-of-the-art ensemble baselines, at the same time can detect a wide range of adversarial examples with a nearly perfect accuracy. Our code is available at: https://github.com/tuananhbui89/Crossing-Collaborative-Ensemble.

Anh Bui, Khanh Doan, Trung Le et al.

Diffusion models have demonstrated remarkable capability in generating high-quality visual content from textual descriptions. However, since these models are trained on large-scale internet data, they inevitably learn undesirable concepts, such as sensitive content, copyrighted material, and harmful or unethical elements. While previous works focus on permanently removing such concepts, this approach is often impractical, as it can degrade model performance and lead to irreversible loss of information. In this work, we introduce a novel concept-hiding approach that makes unwanted concepts inaccessible to public users while allowing controlled recovery when needed. Instead of erasing knowledge from the model entirely, we incorporate a learnable prompt into the cross-attention module, acting as a secure memory that suppresses the generation of hidden concepts unless a secret key is provided. This enables flexible access control -- ensuring that undesirable content cannot be easily generated while preserving the option to reinstate it under restricted conditions. Our method introduces a new paradigm where concept suppression and controlled recovery coexist, which was not feasible in prior works. We validate its effectiveness on the Stable Diffusion model, demonstrating that hiding concepts mitigate the risks of permanent removal while maintaining the model's overall capability.

Sean Lamont, Michael Norrish, Amir Dezfouli et al.

Artificial Intelligence for Theorem Proving has given rise to a plethora of benchmarks and methodologies, particularly in Interactive Theorem Proving (ITP). Research in the area is fragmented, with a diverse set of approaches being spread across several ITP systems. This presents a significant challenge to the comparison of methods, which are often complex and difficult to replicate. Addressing this, we present BAIT, a framework for fair and streamlined comparison of learning approaches in ITP. We demonstrate BAIT's capabilities with an in-depth comparison, across several ITP benchmarks, of state-of-the-art architectures applied to the problem of formula embedding. We find that Structure Aware Transformers perform particularly well, improving on techniques associated with the original problem sets. BAIT also allows us to assess the end-to-end proving performance of systems built on interactive environments. This unified perspective reveals a novel end-to-end system that improves on prior work. We also provide a qualitative analysis, illustrating that improved performance is associated with more semantically-aware embeddings. By streamlining the implementation and comparison of Machine Learning algorithms in the ITP context, we anticipate BAIT will be a springboard for future research.

Shijie Liu, Andrew C. Cullen, Paul Montague et al.

Similar to other machine learning frameworks, Offline Reinforcement Learning (RL) is shown to be vulnerable to poisoning attacks, due to its reliance on externally sourced datasets, a vulnerability that is exacerbated by its sequential nature. To mitigate the risks posed by RL poisoning, we extend certified defenses to provide larger guarantees against adversarial manipulation, ensuring robustness for both per-state actions, and the overall expected cumulative reward. Our approach leverages properties of Differential Privacy, in a manner that allows this work to span both continuous and discrete spaces, as well as stochastic and deterministic environments -- significantly expanding the scope and applicability of achievable guarantees. Empirical evaluations demonstrate that our approach ensures the performance drops to no more than $50\%$ with up to $7\%$ of the training data poisoned, significantly improving over the $0.008\%$ in prior work~\citep{wu_copa_2022}, while producing certified radii that is $5$ times larger as well. This highlights the potential of our framework to enhance safety and reliability in offline RL.

Quoc Viet Vo, Tashreque M. Haq, Paul Montague et al.

Certified defenses promise provable robustness guarantees. We study the malicious exploitation of probabilistic certification frameworks to better understand the limits of guarantee provisions. Now, the objective is to not only mislead a classifier, but also manipulate the certification process to generate a robustness guarantee for an adversarial input certificate spoofing. A recent study in ICLR demonstrated that crafting large perturbations can shift inputs far into regions capable of generating a certificate for an incorrect class. Our study investigates if perturbations needed to cause a misclassification and yet coax a certified model into issuing a deceptive, large robustness radius for a target class can still be made small and imperceptible. We explore the idea of region-focused adversarial examples to craft imperceptible perturbations, spoof certificates and achieve certification radii larger than the source class ghost certificates. Extensive evaluations with the ImageNet demonstrate the ability to effectively bypass state-of-the-art certified defenses such as Densepure. Our work underscores the need to better understand the limits of robustness certification methods.

Andrew C. Cullen, Paul Montague, Sarah M. Erfani et al.

While certified robustness is widely promoted as a solution to adversarial examples in Artificial Intelligence systems, significant challenges remain before these techniques can be meaningfully deployed in real-world applications. We identify critical gaps in current research, including the paradox of detection without distinction, the lack of clear criteria for practitioners to evaluate certification schemes, and the potential security risks arising from users' expectations surrounding ``guaranteed" robustness claims. These create an alignment issue between how certifications are presented and perceived, relative to their actual capabilities. This position paper is a call to arms for the certification research community, proposing concrete steps to address these fundamental challenges and advance the field toward practical applicability.

Shijie Liu, Andrew C. Cullen, Paul Montague et al.

The current state-of-the-art backdoor attacks against Reinforcement Learning (RL) rely upon unrealistically permissive access models, that assume the attacker can read (or even write) the victim's policy parameters, observations, or rewards. In this work, we question whether such a strong assumption is required to launch backdoor attacks against RL. To answer this question, we propose the \underline{S}upply-\underline{C}h\underline{a}in \underline{B}ackdoor (SCAB) attack, which targets a common RL workflow: training agents using external agents that are provided separately or embedded within the environment. In contrast to prior works, our attack only relies on legitimate interactions of the RL agent with the supplied agents. Despite this limited access model, by poisoning a mere $3\%$ of training experiences, our attack can successfully activate over $90\%$ of triggered actions, reducing the average episodic return by $80\%$ for the victim. Our novel attack demonstrates that RL attacks are likely to become a reality under untrusted RL training supply-chains.

Rollin Omari, Junae Kim, Paul Montague

In this paper we explore the challenges and strategies for enhancing the robustness of $k$-means clustering algorithms against adversarial manipulations. We evaluate the vulnerability of clustering algorithms to adversarial attacks, emphasising the associated security risks. Our study investigates the impact of incremental attack strength on training, introduces the concept of transferability between supervised and unsupervised models, and highlights the sensitivity of unsupervised models to sample distributions. We additionally introduce and evaluate an adversarial training method that improves testing performance in adversarial scenarios, and we highlight the importance of various parameters in the proposed training method, such as continuous learning, centroid initialisation, and adversarial step-count.

Anh Bui, Trung Le, He Zhao et al.

Contrastive learning (CL) has recently emerged as an effective approach to learning representation in a range of downstream tasks. Central to this approach is the selection of positive (similar) and negative (dissimilar) sets to provide the model the opportunity to `contrast' between data and class representation in the latent space. In this paper, we investigate CL for improving model robustness using adversarial samples. We first designed and performed a comprehensive study to understand how adversarial vulnerability behaves in the latent space. Based on this empirical evidence, we propose an effective and efficient supervised contrastive learning to achieve model robustness against adversarial attacks. Moreover, we propose a new sample selection strategy that optimizes the positive/negative sets by removing redundancy and improving correlation with the anchor. Extensive experiments show that our Adversarial Supervised Contrastive Learning (ASCL) approach achieves comparable performance with the state-of-the-art defenses while significantly outperforms other CL-based defense methods by using only $42.8\%$ positives and $6.3\%$ negatives.

He Zhao, Thanh Nguyen, Trung Le et al.

Deep neural network image classifiers are reported to be susceptible to adversarial evasion attacks, which use carefully crafted images created to mislead a classifier. Many adversarial attacks belong to the category of dense attacks, which generate adversarial examples by perturbing all the pixels of a natural image. To generate sparse perturbations, sparse attacks have been recently developed, which are usually independent attacks derived by modifying a dense attack's algorithm with sparsity regularisations, resulting in reduced attack efficiency. In this paper, we aim to tackle this task from a different perspective. We select the most effective perturbations from the ones generated from a dense attack, based on the fact we find that a considerable amount of the perturbations on an image generated by dense attacks may contribute little to attacking a classifier. Accordingly, we propose a probabilistic post-hoc framework that refines given dense attacks by significantly reducing the number of perturbed pixels but keeping their attack power, trained with mutual information maximisation. Given an arbitrary dense attack, the proposed model enjoys appealing compatibility for making its adversarial images more realistic and less detectable with fewer perturbations. Moreover, our framework performs adversarial attacks much faster than existing sparse attacks.

Anh Bui, Trung Le, He Zhao et al.

The fact that deep neural networks are susceptible to crafted perturbations severely impacts the use of deep learning in certain domains of application. Among many developed defense models against such attacks, adversarial training emerges as the most successful method that consistently resists a wide range of attacks. In this work, based on an observation from a previous study that the representations of a clean data example and its adversarial examples become more divergent in higher layers of a deep neural net, we propose the Adversary Divergence Reduction Network which enforces local/global compactness and the clustering assumption over an intermediate layer of a deep neural network. We conduct comprehensive experiments to understand the isolating behavior of each component (i.e., local/global compactness and the clustering assumption) and compare our proposed model with state-of-the-art adversarial training methods. The experimental results demonstrate that augmenting adversarial training with our proposed components can further improve the robustness of the network, leading to higher unperturbed and adversarial predictive performances.

He Zhao, Trung Le, Paul Montague et al.

Deep neural network image classifiers are reported to be susceptible to adversarial evasion attacks, which use carefully crafted images created to mislead a classifier. Recently, various kinds of adversarial attack methods have been proposed, most of which focus on adding small perturbations to input images. Despite the success of existing approaches, the way to generate realistic adversarial images with small perturbations remains a challenging problem. In this paper, we aim to address this problem by proposing a novel adversarial method, which generates adversarial examples by imposing not only perturbations but also spatial distortions on input images, including scaling, rotation, shear, and translation. As humans are less susceptible to small spatial distortions, the proposed approach can produce visually more realistic attacks with smaller perturbations, able to deceive classifiers without affecting human predictions. We learn our method by amortized techniques with neural networks and generate adversarial examples efficiently by a forward pass of the networks. Extensive experiments on attacking different types of non-robustified classifiers and robust classifiers with defence show that our method has state-of-the-art performance in comparison with advanced attack parallels.

Yi Han, David Hubczenko, Paul Montague et al.

Recent studies have demonstrated that reinforcement learning (RL) agents are susceptible to adversarial manipulation, similar to vulnerabilities previously demonstrated in the supervised learning setting. While most existing work studies the problem in the context of computer vision or console games, this paper focuses on reinforcement learning in autonomous cyber defence under partial observability. We demonstrate that under the black-box setting, where the attacker has no direct access to the target RL model, causative attacks---attacks that target the training process---can poison RL agents even if the attacker only has partial observability of the environment. In addition, we propose an inversion defence method that aims to apply the opposite perturbation to that which an attacker might use to generate their adversarial samples. Our experimental results illustrate that the countermeasure can effectively reduce the impact of the causative attack, while not significantly affecting the training process in non-attack scenarios.

Yi Han, Benjamin I. P. Rubinstein, Tamas Abraham et al.

Despite the successful application of machine learning (ML) in a wide range of domains, adaptability---the very property that makes machine learning desirable---can be exploited by adversaries to contaminate training and evade classification. In this paper, we investigate the feasibility of applying a specific class of machine learning algorithms, namely, reinforcement learning (RL) algorithms, for autonomous cyber defence in software-defined networking (SDN). In particular, we focus on how an RL agent reacts towards different forms of causative attacks that poison its training process, including indiscriminate and targeted, white-box and black-box attacks. In addition, we also study the impact of the attack timing, and explore potential countermeasures such as adversarial training.