This work introduces a verification framework that provides both sound and complete guarantees for data poisoning attacks during neural network training. We formulate adversarial data manipulation, model training, and test-time evaluation in a single mixed-integer quadratic programming (MIQCP) problem. Finding the global optimum of the proposed formulation provably yields worst-case poisoning attacks, while simultaneously bounding the effectiveness of all possible attacks on the given training pipeline. Our framework encodes both the gradient-based training dynamics and model evaluation at test time, enabling the first exact certification of training-time robustness. Experimental evaluation on small models confirms that our approach delivers a complete characterization of robustness against data poisoning.

As AI-generated images become increasingly photorealistic, distinguishing them from natural images poses a growing challenge. This paper presents a robust detection framework that leverages multiple uncertainty measures to decide whether to trust or reject a model's predictions. We focus on three complementary techniques: Fisher Information, which captures the sensitivity of model parameters to input variations; entropy-based uncertainty from Monte Carlo Dropout, which reflects predictive variability; and predictive variance from a Deep Kernel Learning framework using a Gaussian Process classifier. To integrate these diverse uncertainty signals, Particle Swarm Optimisation is used to learn optimal weightings and determine an adaptive rejection threshold. The model is trained on Stable Diffusion-generated images and evaluated on GLIDE, VQDM, Midjourney, BigGAN, and StyleGAN3, each introducing significant distribution shifts. While standard metrics such as prediction probability and Fisher-based measures perform well in distribution, their effectiveness degrades under shift. In contrast, the Combined Uncertainty measure consistently achieves an incorrect rejection rate of approximately 70 percent on unseen generators, successfully filtering most misclassified AI samples. Although the system occasionally rejects correct predictions from newer generators, this conservative behaviour is acceptable, as rejected samples can support retraining. The framework maintains high acceptance of accurate predictions for natural images and in-domain AI data. Under adversarial attacks using FGSM and PGD, the Combined Uncertainty method rejects around 61 percent of successful attacks, while GP-based uncertainty alone achieves up to 80 percent. Overall, the results demonstrate that multi-source uncertainty fusion provides a resilient and adaptive solution for AI-generated image detection.

The impact of inference-time data perturbation (e.g., adversarial attacks) has been extensively studied in machine learning, leading to well-established certification techniques for adversarial robustness. In contrast, certifying models against training data perturbations remains a relatively under-explored area. These perturbations can arise in three critical contexts: adversarial data poisoning, where an adversary manipulates training samples to corrupt model performance; machine unlearning, which requires certifying model behavior under the removal of specific training data; and differential privacy, where guarantees must be given with respect to substituting individual data points. This work introduces Abstract Gradient Training (AGT), a unified framework for certifying robustness of a given model and training procedure to training data perturbations, including bounded perturbations, the removal of data points, and the addition of new samples. By bounding the reachable set of parameters, i.e., establishing provable parameter-space bounds, AGT provides a formal approach to analyzing the behavior of models trained via first-order optimization methods.

The continually advancing quality of deepfake technology exacerbates the threats of disinformation, fraud, and harassment by making maliciously-generated synthetic content increasingly difficult to distinguish from reality. We introduce a simple yet effective two-stage detection method that achieves an AUROC of over 99.8% on contemporary deepfakes. However, this high performance is short-lived. We show that models trained on this data suffer a recall drop of over 30% when evaluated on deepfakes created with generation techniques from just six months later, demonstrating significant decay as threats evolve. Our analysis reveals two key insights for robust detection. Firstly, continued performance requires the ongoing curation of large, diverse datasets. Second, predictive power comes primarily from static, frame-level artifacts, not temporal inconsistencies. The future of effective deepfake detection therefore depends on rapid data collection and the development of advanced frame-level feature detectors.

Federated learning (FL) is a privacy-preserving machine learning technique that facilitates collaboration among participants across demographics. FL enables model sharing, while restricting the movement of data. Since FL provides participants with independence over their training data, it becomes susceptible to poisoning attacks. Such collaboration also propagates bias among the participants, even unintentionally, due to different data distribution or historical bias present in the data. This paper proposes an intentional fairness attack, where a client maliciously sends a biased model, by increasing the fairness loss while training, even considering homogeneous data distribution. The fairness loss is calculated by solving an optimization problem for fairness metrics such as demographic parity and equalized odds. The attack is insidious and hard to detect, as it maintains global accuracy even after increasing the bias. We evaluate our attack against the state-of-the-art Byzantine-robust and fairness-aware aggregation schemes over different datasets, in various settings. The empirical results demonstrate the attack efficacy by increasing the bias up to 90\%, even in the presence of a single malicious client in the FL system.

Before adopting a new large language model (LLM) architecture, it is critical to understand vulnerabilities accurately. Existing evaluations can be difficult to trust, often drawing conclusions from LLMs that are not meaningfully comparable, relying on heuristic inputs or employing metrics that fail to capture the inherent uncertainty. In this paper, we propose a principled and practical end-to-end framework for evaluating LLM vulnerabilities to prompt injection attacks. First, we propose practical approaches to experimental design, tackling unfair LLM comparisons by considering two practitioner scenarios: when training an LLM and when deploying a pre-trained LLM. Second, we address the analysis of experiments and propose a Bayesian hierarchical model with embedding-space clustering. This model is designed to improve uncertainty quantification in the common scenario that LLM outputs are not deterministic, test prompts are designed imperfectly, and practitioners only have a limited amount of compute to evaluate vulnerabilities. We show the improved inferential capabilities of the model in several prompt injection attack settings. Finally, we demonstrate the pipeline to evaluate the security of Transformer versus Mamba architectures. Our findings show that consideration of output variability can suggest less definitive findings. However, for some attacks, we find notably increased Transformer and Mamba-variant vulnerabilities across LLMs with the same training data or mathematical ability.

With the growing accessibility and wide adoption of large language models, concerns about their safety and alignment with human values have become paramount. In this paper, we identify a concerning phenomenon: Reasoning-Induced Misalignment (RIM), in which misalignment emerges when reasoning capabilities strengthened-particularly when specific types of reasoning patterns are introduced during inference or training. Beyond reporting this vulnerability, we provide the first mechanistic account of its origins. Through representation analysis, we discover that specific attention heads facilitate refusal by reducing their attention to CoT tokens, a mechanism that modulates the model's rationalization process during inference. During training, we find significantly higher activation entanglement between reasoning and safety in safety-critical neurons than in control neurons, particularly after fine-tuning with those identified reasoning patterns. This entanglement strongly correlates with catastrophic forgetting, providing a neuron-level explanation for RIM.