Ivan Sabolić

Supervised fine-tuning is the predominant approach for adapting autoregressive vision-language models to downstream tasks. Recent work has shown that this paradigm is highly vulnerable to backdoor attacks, and that existing defenses are ineffective in open-ended generation settings. In response, we propose BYORn, a backdoor-robust fine-tuning framework motivated by the observation that poisoned target responses are often semantically implausible given the corresponding image-text inputs and a pretrained model. BYORn identifies such misaligned responses and dynamically replaces them with alternative responses generated by the model, thereby breaking the correlation between triggers and target outputs. The resulting objective gradient corresponds to the gradient of the empirical estimate of the population risk upper bound over the clean data distribution. Empirically, BYORn consistently improves robustness to backdoor attacks while preserving clean-task performance, establishing a new trade-off frontier between generalization and attack success rate. Finally, we demonstrate that BYORn remains effective against adaptive attacks specifically designed to circumvent the proposed defense.

We propose VIBE, a model-agnostic framework that trains classifiers resilient to backdoor attacks. The key concept behind our approach is to treat malicious inputs and corrupted labels from the training dataset as observed random variables, while the actual clean labels are latent. VIBE then recovers the corresponding latent clean label posterior through variational inference. The resulting training procedure follows the expectation-maximization (EM) algorithm. The E-step infers the clean pseudolabels by solving an entropy-regularized optimal transport problem, while the M-step updates the classifier parameters via gradient descent. Being modular, VIBE can seamlessly integrate with recent advancements in self-supervised representation learning, which enhance its ability to resist backdoor attacks. We experimentally validate the method effectiveness against contemporary backdoor attacks on standard datasets, a large-scale setup with 1k classes, and a dataset poisoned with multiple attacks. VIBE consistently outperforms previous defenses across all tested scenarios.

Backdoor attacks change a small portion of training data by introducing hand-crafted triggers and rewiring the corresponding labels towards a desired target class. Training on such data injects a backdoor which causes malicious inference in selected test samples. Most defenses mitigate such attacks through various modifications of the discriminative learning procedure. In contrast, this paper explores an approach based on generative modelling of per-class distributions in a self-supervised representation space. Interestingly, these representations get either preserved or heavily disturbed under recent backdoor attacks. In both cases, we find that per-class generative models allow to detect poisoned data and cleanse the dataset. Experiments show that training on cleansed dataset greatly reduces the attack success rate and retains the accuracy on benign inputs.