The research landscape in the domain of backdoor attacks against machine learning models is rapidly evolving, with a strong focus on developing more sophisticated, resilient, and stealthy attack methodologies. Recent advancements highlight the vulnerability of various model architectures, including vision transformers, code generation models, and no-reference image quality assessment models, to backdoor attacks. Innovations in trigger design, such as leveraging attention mechanisms and discrete cosine transform domains, are enhancing the effectiveness and stealthiness of these attacks. Additionally, there is a growing emphasis on model-agnostic and data-free approaches, which do not require access to clean data or modifications to the model architecture, making them more practical and stealthy. The robustness of these attacks against state-of-the-art defenses is also being rigorously tested, with promising results indicating the need for more proactive defense strategies. Notably, the integration of prompt tuning in vision-language models for detecting unseen backdoored images represents a significant step forward in proactive adversarial defense, offering a new paradigm for identifying and mitigating backdoor threats. Overall, the field is moving towards more adaptive, universal, and resilient attack strategies, necessitating equally innovative defense mechanisms to maintain the integrity and reliability of machine learning models in critical applications.
Evolving Backdoor Attacks and Proactive Defense Strategies
Sources
Megatron: Evasive Clean-Label Backdoor Attacks against Vision Transformer
Backdooring Outlier Detection Methods: A Novel Attack Approach
SABER: Model-agnostic Backdoor Attack on Chain-of-Thought in Neural Code Generation
An Effective and Resilient Backdoor Attack Framework against Deep Neural Networks and Vision Transformers
Data Free Backdoor Attacks
Backdoor Attacks against No-Reference Image Quality Assessment Models via A Scalable Trigger
Impact of Sampling Techniques and Data Leakage on XGBoost Performance in Credit Card Fraud Detection
Stealthy and Robust Backdoor Attack against 3D Point Clouds through Additional Point Features
Doubly-Universal Adversarial Perturbations: Deceiving Vision-Language Models Across Both Images and Text with a Single Perturbation
Backdoor attacks on DNN and GBDT -- A Case Study from the insurance domain
Proactive Adversarial Defense: Harnessing Prompt Tuning in Vision-Language Models to Detect Unseen Backdoored Images