Towards a Robust Defense: A Multifaceted Approach to the Detection and Mitigation of Neural Backdoor Attacks through Feature Space Exploration and Analysis

Liuwan Zhu

Abstract: From voice assistants to self-driving vehicles, machine learning(ML), especially deep learning, revolutionizes the way we work and live, through the wide adoption in a broad range of applications. Unfortunately, this widespread use makes deep learning-based systems a desirable target for cyberattacks, such as generating adversarial examples to fool a deep learning system to make wrong decisions. In particular, many recent studies have revealed that attackers can corrupt the training of a deep learning model, e.g., through data poisoning, or distribute a deep learning model they created with “backdoors” planted, e.g., distributed as part of a software library, so that the attacker can easily craft system inputs that grant unauthorized access or lead to catastrophic errors or failures. This dissertation aims to develop a multifaceted approach for detecting and mitigating such neural backdoor attacks by exploiting their unique characteristics in the feature space. First of all, a framework called GangSweep is designed to utilize the capabilities of Generative Adversarial Networks (GAN) to approximate poisoned sample distributions in the feature space, to detect neural backdoor attacks. Unlike conventional methods, GangSweep exposes all attacker-induced artifacts, irrespective of their complexity or obscurity. By leveraging the statistical disparities between these artifacts and natural adversarial perturbations, an efficient detection scheme is devised. Accordingly, the backdoored model can be purified through label correction and fine-tuning Secondly, this dissertation focuses on the sample-targeted backdoor attacks, a variant of neural backdoor that targets specific samples. Given the absence of explicit triggers in such models, traditional detection methods falter. Through extensive analysis, I have identified a unique feature space property of these attacks, where they induce boundary alterations, creating discernible “pockets” around target samples. Based on this critical observation, I introduce a novel defense scheme that encapsulates these malicious pockets within a tight convex hull in the feature space, and then design an algorithm to identify such hulls and remove the backdoor through model fine-tuning. The algorithm demonstrates high efficacy against a spectrum of sample-targeted backdoor attacks. Lastly, I address the emerging challenge of backdoor attacks in multimodal deep neural networks, in particular vision-language model, a growing concern in real-world applications. Discovering that there is a strong association between the image trigger and the target text in the feature space of the backdoored vision-language model, I design an effective algorithm to expose the malicious text and image trigger by jointly searching in the shared feature space of the vision and language modalities.