This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
Encoders pre-trained using self-supervised learning (SSL) are vulnerable to backdoor attacks. This paper proposes MIMIC, a mutual information-based backdoor mitigation technique to mitigate backdoor attacks on pre-trained encoders. MIMIC treats a potentially backdoored encoder as a teacher network and uses knowledge distillation to distill clean student encoders from the teacher network. MIMIC performs distillation by leveraging mutual information between each layer and extracted features to identify the locations of positive knowledge in the teacher network.
Takeaways, Limitations
•
MIMIC effectively mitigates backdoor attacks on pre-trained encoders.
•
MIMIC outperforms existing techniques using a small amount of clean data.
•
This paper developed a backdoor mitigation technique using mutual information.
•
Limitations in the paper was not specifically mentioned.
