An Explainable AI-based Complementary Attention Mechanism for Detecting Identity Swaps

Supriyo Sadhya, Xiaojun Qi

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: deep learning, fake content, fake faces, identity swap, scaled spatial attention, layer-integrated channel attention, LIME, deepfake, faceswap

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Deep learning techniques have quickly led to the generation of a large number of realistic fake content by accessing large-scale publicly available databases. The emergence of deepfake technology has given rise to concerns related to the creation and dissemination of manipulated multimedia content because of its use in social media to generate fake news. One prevalent application of this technology is identity swap, wherein faces are exchanged within images and videos to create convincing yet fabricated visual narratives. Thus, the detection of identity swaps has become an increasingly important research area in the field of digital forensics. This paper presents a complementary attention-based deep learning system for the detection of identity swaps. Specifically, it incorporates our proposed simple Layer-Integrated Channel Attention (LICA) and Scaled Spatial Attention (SSA) mechanisms in the VGG network architecture to respectively capture the importance along each channel and at each spatial location to distinguish real faces from manipulated faces. It further incorporates Local Interpretable Model-agnostic Explanations (LIME) as the explainable AI technique to provide a more in-depth transparent analysis of its effectiveness towards improved detection performance. Our extensive experimental results demonstrate that the proposed system outperforms state-of-the-art systems in terms of accuracy and area under curve metrics in detecting fake faces generated by identity swaps. The LIME further provides a deeper understanding of the decision-making process and facilitates trust and accountability by combining the power of CNNs with the transparency of explainable AI.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8919