Alireza Dehghanpour Farashah

Mahdi Salmani, Alireza Dehghanpour Farashah, Mohammad Azizmalayeri et al.

Despite the remarkable success achieved by deep learning algorithms in various domains, such as computer vision, they remain vulnerable to adversarial perturbations. Adversarial Training (AT) stands out as one of the most effective solutions to address this issue; however, single-step AT can lead to Catastrophic Overfitting (CO). This scenario occurs when the adversarially trained network suddenly loses robustness against multi-step attacks like Projected Gradient Descent (PGD). Although several approaches have been proposed to address this problem in Convolutional Neural Networks (CNNs), we found out that they do not perform well when applied to Vision Transformers (ViTs). In this paper, we propose Blacksmith, a novel training strategy to overcome the CO problem, specifically in ViTs. Our approach utilizes either of PGD-2 or Fast Gradient Sign Method (FGSM) randomly in a mini-batch during the adversarial training of the neural network. This will increase the diversity of our training attacks, which could potentially mitigate the CO issue. To manage the increased training time resulting from this combination, we craft the PGD-2 attack based on only the first half of the layers, while FGSM is applied end-to-end. Through our experiments, we demonstrate that our novel method effectively prevents CO, achieves PGD-2 level performance, and outperforms other existing techniques including N-FGSM, which is the state-of-the-art method in fast training for CNNs.

'Mina Arzaghi', 'Alireza Dehghanpour Farashah', 'Florian Carichon' et al.

Large Language Models (LLMs) exhibit socio-economic biases that can propagate into downstream tasks. While prior studies have questioned whether intrinsic bias in LLMs affects fairness at the downstream task level, this work empirically investigates the connection. We present a unified evaluation framework to compare intrinsic bias mitigation via concept unlearning with extrinsic bias mitigation via counterfactual data augmentation (CDA). We examine this relationship through real-world financial classification tasks, including salary prediction, employment status, and creditworthiness assessment. Using three open-source LLMs, we evaluate models both as frozen embedding extractors and as fine-tuned classifiers. Our results show that intrinsic bias mitigation through unlearning reduces intrinsic gender bias by up to 94.9%, while also improving downstream task fairness metrics, such as demographic parity by up to 82%, without compromising accuracy. Our framework offers practical guidance on where mitigation efforts can be most effective and highlights the importance of applying early-stage mitigation before downstream deployment.