Abstract
The security of AI models poses significant challenges, as sensitive user information can potentially be inferred from the models, leading to privacy breaches. To address this, machine unlearning methods aim to remove specific data from a trained model, effectively eliminating the training traces of those data. However, most existing approaches focus primarily on supervised learning scenarios, leaving the unlearning of contrastive learning models underexplored. This paper proposes a novel fine-tuning-based unlearning method tailored for contrastive learning models. The approach introduces a third-party dataset to ensure that the model’s outputs for forgotten data align with those of the third-party dataset, thereby removing identifiable training traces. A comprehensive loss function is designed, encompassing three objectives: preserving model accuracy, constraining gradients to make forgotten and third-party data indistinguishable, and reducing model confidence on the third-party dataset. The experimental results demonstrate the effectiveness of the proposed method. Membership inference attacks conducted before and after unlearning show that the forgotten data’s prediction distribution becomes indistinguishable from that of the third-party data, validating the success of the unlearning process. Moreover, the proposed method achieves this with minimal performance degradation, making it suitable for practical applications in privacy-preserving AI.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Similar Papers
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.