Artificial Intelligence (AI) models are increasingly used in complex systems such as autonomous vehicles (AVs), where safety and explainability are critical. However, existing explainable AI (xAI) methods focus on model-level transparency while neglecting system-level safety explanations (Gap 1), and prior applications of large language models (LLMs) in AVs often view the AV as a whole, overlooking potential risks arising from interactions among its internal components (Gap 2). To address these gaps, we propose a framework that integrates LLMs with System Theoretic Process Analysis (STPA), a structured method to analyse hazards and assess safety, to improve AV safety assurance. Our framework leverages LLMs for scenario analysis while incorporating STPA to identify unsafe control actions (UCAs) and filter them with real-world video data. We evaluated our method against Lingo-2 (a vision-language-action model developed by Wayve) in a simulated environment, demonstrating superior STPA-based explanations. To evaluate the framework, we employed two ground truth references for accuracy verification and conducted robustness testing, which outperforming traditional LLM-based explainers, as also confirmed by expert evaluations.

Text-to-Image (T2I) Diffusion Models (DMs) have garnered widespread attention for their impressive advancements in image generation. However, their growing popularity has raised ethical and social concerns related to key non-functional properties of trustworthiness, such as robustness, fairness, security, privacy, and explainability, similar to those in traditional deep learning (DL) tasks. Conventional approaches for studying trustworthiness in DL tasks often inadequate for T2I DMs because of their unique characteristics, e.g., multi-modal nature, stochastic generation process, and high computational cost. Given these challenge, recent efforts have been made to develop new methods for investigating trustworthiness in T2I DMs via various means, including falsification, enhancement, verification and assessment. However, there is a notable lack of in-depth analysis concerning those non-functional properties and means. In this survey, we provide a timely and focused review of the literature on trustworthy T2I DMs, covering a concise-structured taxonomy from the perspectives of property, means, benchmarks and applications. Our review begins with an introduction to essential preliminaries of T2I DMs, and then we summarize key definitions and metrics specific to T2I tasks, based on which we analyze the corresponding means in recent literature. Additionally, we review benchmarks and domain applications of T2I DMs. Finally, we highlight the gaps in current research, discuss the limitations of existing methods, and propose future research directions to advance the development of trustworthy T2I DMs. Furthermore, we keep up-to-date updates in this field to track the latest developments and maintain our GitHub repository at: https://github.com/wellzline/Trustworthy_T2I_DMs.

Personalized AI applications such as DreamBooth enable the generation of customized content from user images, but also raise significant privacy concerns, particularly the risk of facial identity leakage. Recent defense mechanisms like Anti-DreamBooth attempt to mitigate this risk by injecting adversarial perturbations into user photos to prevent successful personalization. However, we identify two critical yet overlooked limitations of these methods. First, the adversarial examples often exhibit perceptible artifacts such as conspicuous patterns or stripes, making them easily detectable as manipulated content. Second, the perturbations are highly fragile, as even a simple, non-learned filter can effectively remove them, thereby restoring the model's ability to memorize and reproduce user identity. To investigate this vulnerability, we propose a novel evaluation framework, AntiDB_Purify, to systematically evaluate existing defenses under realistic purification threats, including both traditional image filters and adversarial purification. Results reveal that none of the current methods maintains their protective effectiveness under such threats. These findings highlight that current defenses offer a false sense of security and underscore the urgent need for more imperceptible and robust protections to safeguard user identity in personalized generation.

Deep learning (DL) has shown transformative potential across industries, yet its sensitivity to adversarial examples (AEs) limits its reliability and broader deployment. Research on $D L$ robustness has developed various techniques, with adversarial training (AT) established as a leading approach to counter AEs. Traditional AT focuses on worstcase robustness (WCR), but recent work has introduced probabilistic robustness (PR), which evaluates the likelihood of AEs within a local perturbation range, providing an overall assessment of the model's robustness and acknowledging residual risks that are more practical to manage. However, existing AT methods are fundamentally designed to improve WCR, and no dedicated methods currently target PR. To bridge this gap, we formulate a new min-max optimization as the theoretical foundation for PR-focused $A T$, and introduce an AT-PR training scheme with numerical algorithms to solve the new optimization problem. Our experiments, based on 70 DL models trained on common datasets and diverse architectures, demonstrate that: i) $A T-P R$ achieves higher improvements in PR than AT-WCR methods; ii) it shows more consistent effectiveness across varying local inputs; iii) it exhibits a reduced trade-off in model's generalization.

Out-of-distribution (OOD) detectors play a vital role in distinguishing between OOD data and in-distribution data. However, the vulnerability of OOD detectors to natural perturbations, such as rotation and lighting variations, can potentially lead to catastrophic accidents in safety-critical applications. Current attack techniques lack robustness guarantees for OOD detectors. Neural network (NN) verification methods are limited to standard NN structures and cannot be applied to OOD detectors due to their non-standard structure. To address this issue, we propose a verification framework called Vood that offers robustness guarantees for OOD detectors under natural perturbations. Our approach begins by proving the Lipschitz continuity of most OOD detection functions under natural transformations. We then estimate the Lipschitz constant using Extreme Value Theory, incorporating a dynamically estimated safety factor. Vood transforms the verification problem into an optimization challenge, which is then effectively addressed using space-filling Lipschitz optimization techniques. Additionally, Vood is a black-box verifier, which can tackle natural perturbations on a wide range of OOD detectors. Through empirical analysis, we demonstrate that Vood outperforms baseline methods in both accuracy and efficiency. Our work represents a pioneering effort in establishing robustness verification for OOD detectors with provable guarantees.

Recent advances in large vision-language models (VLMs) have demonstrated remarkable success across a wide range of visual understanding tasks. However, the robustness of these models against jailbreak attacks remains an open challenge. In this work, we propose a universal certified defence framework to safeguard VLMs rigorously against potential visual jailbreak attacks. First, we proposed a novel distance metric to quantify semantic discrepancies between malicious and intended responses, capturing subtle differences often overlooked by conventional cosine similarity-based measures. Then, we devise a regressed certification approach that employs randomized smoothing to provide formal robustness guarantees against both adversarial and structural perturbations, even under black-box settings. Complementing this, our feature-space defence introduces noise distributions (e.g., Gaussian, Laplacian) into the latent embeddings to safeguard against both pixel-level and structure-level perturbations. Our results highlight the potential of a formally grounded, integrated strategy toward building more resilient and trustworthy VLMs.

Large language models have been widely applied, but can inadvertently encode sensitive or harmful information, raising significant safety concerns. Machine unlearning has emerged to alleviate this concern; however, existing training-time unlearning approaches, relying on coarse-grained loss combinations, have limitations in precisely separating knowledge and balancing removal effectiveness with model utility. In contrast, we propose Fine-grained Activation manipuLation by Contrastive Orthogonal uNalignment (FALCON), a novel representation-guided unlearning approach that leverages information-theoretic guidance for efficient parameter selection, employs contrastive mechanisms to enhance representation separation, and projects conflict gradients onto orthogonal subspaces to resolve conflicts between forgetting and retention objectives. Extensive experiments demonstrate that FALCON achieves superior unlearning effectiveness while maintaining model utility, exhibiting robust resistance against knowledge recovery attempts.

Deep learning models are notoriously vulnerable to imperceptible perturbations. Most existing research centers on adversarial robustness (AR), which evaluates models under worst-case scenarios by examining the existence of deterministic adversarial examples (AEs). In contrast, probabilistic robustness (PR) adopts a statistical perspective, measuring the probability that predictions remain correct under stochastic perturbations. While PR is widely regarded as a practical complement to AR, dedicated training methods for improving PR are still relatively underexplored, albeit with emerging progress. Among the few PR-targeted training methods, we identify three limitations: i non-comparable evaluation protocols; ii limited comparisons to strong AT baselines despite anecdotal PR gains from AT; and iii no unified framework to compare the generalization of these methods. Thus, we introduce PRBench, the first benchmark dedicated to evaluating improvements in PR achieved by different robustness training methods. PRBench empirically compares most common AT and PR-targeted training methods using a comprehensive set of metrics, including clean accuracy, PR and AR performance, training efficiency, and generalization error (GE). We also provide theoretical analysis on the GE of PR performance across different training methods. Main findings revealed by PRBench include: AT methods are more versatile than PR-targeted training methods in terms of improving both AR and PR performance across diverse hyperparameter settings, while PR-targeted training methods consistently yield lower GE and higher clean accuracy. A leaderboard comprising 229 trained models across 7 datasets and 10 model architectures is publicly available at https://wellzline.github.io/PRBenchLeaderboard/.

Camera-based Bird's Eye View (BEV) perception models receive increasing attention for their crucial role in autonomous driving, a domain where concerns about the robustness and reliability of deep learning have been raised. While only a few works have investigated the effects of randomly generated semantic perturbations, aka natural corruptions, on the multi-view BEV detection task, we develop a black-box robustness evaluation framework that adversarially optimises three common semantic perturbations: geometric transformation, colour shifting, and motion blur, to deceive BEV models, serving as the first approach in this emerging field. To address the challenge posed by optimising the semantic perturbation, we design a smoothed, distance-based surrogate function to replace the mAP metric and introduce SimpleDIRECT, a deterministic optimisation algorithm that utilises observed slopes to guide the optimisation process. By comparing with randomised perturbation and two optimisation baselines, we demonstrate the effectiveness of the proposed framework. Additionally, we provide a benchmark on the semantic robustness of ten recent BEV models. The results reveal that PolarFormer, which emphasises geometric information from multi-view images, exhibits the highest robustness, whereas BEVDet is fully compromised, with its precision reduced to zero.

Vision Language Models (VLMs) have demonstrated impressive inference capabilities, but remain vulnerable to jailbreak attacks that can induce harmful or unethical responses. Existing defence methods are predominantly white-box approaches that require access to model parameters and extensive modifications, making them costly and impractical for many real-world scenarios. Although some black-box defences have been proposed, they often impose input constraints or require multiple queries, limiting their effectiveness in safety-critical tasks such as autonomous driving. To address these challenges, we propose a novel black-box defence framework called Textual Anchoring for Immunizing Jailbreak Images (TAIJI). TAIJI leverages key phrase-based textual anchoring to enhance the model's ability to assess and mitigate the harmful content embedded within both visual and textual prompts. Unlike existing methods, TAIJI operates effectively with a single query during inference, while preserving the VLM's performance on benign tasks. Extensive experiments demonstrate that TAIJI significantly enhances the safety and reliability of VLMs, providing a practical and efficient solution for real-world deployment.

Deep learning models are intrinsically susceptible to textual adversarial attacks on social media, where the perturbed text can trigger aberrant behaviours of victim models and threaten security and privacy. In this paper, we present a novel word-level attack called SCALA: a Synonym-based desCending And repLace-back Ascending mechanism. Our focus is on the efficient production of adversarial examples, with a particular emphasis on minimizing human perceptibility while ensuring the visual resemblance and semantic correctness. The merits of our attacking solution lie in being: i) imperceptible - it keeps a very low word perturbation rate based on the Hamming ( L-0 -norm) distance, thus achieving heightened deceptiveness validated through human evaluations; ii) efficient - our tensor-based parallelization strategy ensures the attacking efficiency compared with baselines; iii) effective - it surpasses seven state-of-the-art attacks on five target models in terms of reducing after-attack accuracy; iv) practical - black-box score-based setting ensures that the adversary only needs to query target models for confidence scores; and v) transferable - our attack shows competitive transferability on the generated adversarial examples. We release our code SCALA via https://github.com/TrustAI/SCALA

As Large Language Models (LLMs) become more integrated into our daily lives, it is crucial to identify and mitigate their risks, especially when the risks can have profound impacts on human users and societies. Guardrails, which filter the inputs or outputs of LLMs, have emerged as a core safeguarding technology. This position paper takes a deep look at current open-source solutions (Llama Guard, Nvidia NeMo, Guardrails AI), and discusses the challenges and the road towards building more complete solutions. Drawing on robust evidence from previous research, we advocate for a systematic approach to construct guardrails for LLMs, based on comprehensive consideration of diverse contexts across various LLMs applications. We propose employing socio-technical methods through collaboration with a multi-disciplinary team to pinpoint precise technical requirements, exploring advanced neural-symbolic implementations to embrace the complexity of the requirements, and developing verification and testing to ensure the utmost quality of the final product.

Formal methods and machine learning are two research fields with drastically different foundations and philosophies. Formal methods utilise mathematically rigorous techniques for software and hardware systems' specification, development and verification. Machine learning focuses on pragmatic approaches to gradually improve a parameterised model by observing a training data set. While historically, the two fields lack communication, this trend has changed in the past few years with an outburst of research interest in the robustness verification of neural networks. This paper will briefly review these works, and focus on the urgent need for broader and more in-depth communication between the two fields, with the ultimate goal of developing learning-enabled systems with excellent performance and acceptable safety and security. We present a specification language, MLS2, and show that it can express a set of known safety and security properties, including generalisation, uncertainty, robustness, data poisoning, backdoor, model stealing, membership inference, model inversion, interpretability, and fairness. To verify MLS2 properties, we promote the global optimisation-based methods, which have provable guarantees on the convergence to the optimal solution. Many of them have theoretical bounds on the gap between current solutions and the optimal solution.

Adversarial training is extensively utilized to improve the adversarial robustness of deep neural networks. Yet, mitigating the degradation of standard generalization performance in adversarial-trained models remains an open problem. This paper attempts to resolve this issue through the lens of model complexity. First, We leverage the Fisher-Rao norm, a geometrically invariant metric for model complexity, to establish the non-trivial bounds of the Cross-Entropy Loss-based Rademacher complexity for a ReLU-activated Multi-Layer Perceptron. Then we generalize a complexity-related variable, which is sensitive to the changes in model width and the trade-off factors in adversarial training. Moreover, intensive empirical evidence validates that this variable highly correlates with the generalization gap of Cross-Entropy loss between adversarial-trained and standard-trained models, especially during the initial and final phases of the training process. Building upon this observation, we propose a novel regularization framework, called Logit-Oriented Adversarial Training (LOAT), which can mitigate the trade-off between robustness and accuracy while imposing only a negligible increase in computational overhead. Our extensive experiments demonstrate that the proposed regularization strategy can boost the performance of the prevalent adversarial training algorithms, including PGD-AT, TRADES, TRADES (LSE), MART, and DM-AT, across various network architectures. Our code will be available at https://github.com/TrustAI/LOAT.

Language models are vulnerable to clandestinely modified data and manipulation by attackers. Despite considerable research dedicated to enhancing robustness against adversarial attacks, the realm of provable robustness for backdoor attacks remains relatively unexplored. In this paper, we initiate a pioneering investigation into the certified robustness of NLP models against backdoor triggers.We propose a model-agnostic mechanism for large-scale models that applies to complex model structures without the need for assessing model architecture or internal knowledge. More importantly, we take recent advances in randomized smoothing theory and propose a novel weight-based distribution algorithm to enable semantic similarity and provide theoretical robustness guarantees.Experimentally, we demonstrate the efficacy of our approach across a diverse range of datasets and tasks, highlighting its utility in mitigating backdoor triggers. Our results show strong performance in terms of certified accuracy, scalability, and semantic preservation.

Adversarial robustness and privacy of deep learning (DL) models are two widely studied topics in AI security. Adversarial training (AT) is an effective approach to improve the robustness of DL models against adversarial attacks. However, while models with AT demonstrate enhanced robustness, they become more susceptible to membership inference attacks (MIAs), thus increasing the risk of privacy leakage. This indicates a negative trade-off between adversarial robustness and privacy in general deep learning models. Visual prompting is a novel model reprogramming (MR) technique used for fine-tuning pre-trained models, achieving good performance in vision tasks, especially when combined with the label mapping technique. However, the performance of label-mapping-based visual prompting (LM-VP) under adversarial attacks and MIAs lacks evaluation. In this work, we regard the MR of LM-VP as a unified entity, referred to as the LM-VP model, and take a step toward jointly evaluating the adversarial robustness and privacy of LM-VP models. Experimental results show that the choice of pre-trained models significantly affects the white-box adversarial robustness of LM-VP, and standard AT even substantially degrades its performance. In contrast, transfer AT-trained LM-VP achieves a good trade-off between transferred adversarial robustness and privacy, a finding that has been consistently validated across various pre-trained models. Code is available at https://github.com/TrustAI/TARP-VP.

Although adversarial training (AT) has proven effective in enhancing the model's robustness, the recently revealed issue of fairness in robustness has not been well addressed, i.e. the robust accuracy varies significantly among different categories. In this paper, instead of uniformly evaluating the model's average class performance, we delve into the issue of robust fairness, by considering the worst-case distribution across various classes. We propose a novel learning paradigm, named Fairness-Aware Adversarial Learning (FAAL). As a generalization of conventional AT, we re-define the problem of adversarial training as a min-max-max framework, to ensure both robustness and fairness of the trained model. Specifically, by taking advantage of distributional robust optimization, our method aims to find the worst distribution among different categories, and the solution is guaranteed to obtain the upper bound performance with high probability. In particular, FAAL can fine-tune an unfair robust model to be fair within only two epochs, without compromising the overall clean and robust accuracies. Extensive experiments on various image datasets validate the superior performance and efficiency of the proposed FAAL compared to other state-of-the-art methods.

Reinforcement Learning (RL) has achieved remarkable success in safety-critical areas, but it can be weakened by adversarial attacks. Recent studies have introduced "smoothed policies" to enhance its robustness. Yet, it is still challenging to establish a provable guarantee to certify the bound of its total reward. Prior methods relied primarily on computing bounds using Lipschitz continuity or calculating the probability of cumulative reward being above specific thresholds. However, these techniques are only suited for continuous perturbations on the RL agent's observations and are restricted to perturbations bounded by the l(2)-norm. To address these limitations, this paper proposes a general blackbox certification method, called ReCePS, which is capable of directly certifying the cumulative reward of the smoothed policy under various l(p)-norm bounded perturbations. Furthermore, we extend our methodology to certify perturbations on action spaces. Our approach leverages f-divergence to measure the distinction between the original distribution and the perturbed distribution, subsequently determining the certification bound by solving a convex optimisation problem. We provide a comprehensive theoretical analysis and run experiments in multiple environments. Our results show that our method not only improves the tightness of certified lower bound of the mean cumulative reward but also demonstrates better efficiency than state-of-the-art methods.

The implicit bias of gradient descent has long been considered the primary mechanism explaining the superior generalization of over-parameterized neural networks without overfitting, even when the training error is zero. However, the implicit bias toward adversarial robustness has rarely been considered in the research community, although it is crucial for the trustworthiness of machine learning models. To fill this gap, in this paper, we explore whether consecutive layers collaborate to strengthen adversarial robustness during gradient descent. By quantifying this collaboration between layers using our proposed concept, co-correlation, we demonstrate a monotonically increasing trend in co-correlation, which implies a decreasing trend in adversarial robustness during gradient descent. Additionally, we observe different behaviours between narrow and wide neural networks during gradient descent. We conducted extensive experiments that verified our proposed theorems.