Work

ディープフェイク検出におけるバックドア攻撃の分析と防御戦略の改善

Year: 2025
Type: dissertation
Abstract: In the thriving era of information technology, media content has become a significant component of our society.An immense volume of media data is generated, transformed, and transmitted every second, or even millisecond.Concurrently, various methods for manipulating images have been developed and enhanced to cater to the needs of users.Despite the undeniable contributions of technology in this area, there is another critical dimension to consider: security concerns.Deepfake technology, which manipulates facial images, poses a serious threat to facial authentication systems and the integrity of content, thereby undermining the trustworthiness of media content.As a result, deepfake detection has garnered considerable attention from the research community, leading to numerous proposals for detection methods aimed at distinguishing real from fake images.This thesis specifically examines the scenario in which deepfake detection methods are susceptible to backdoor attacks.We conducted a thorough and comprehensive evaluation of backdoor attacks and defense strategies in the context of deepfake detection.Our findings offer insightful analyses of the effectiveness of both attacks and defenses.Furthermore, we proposed a novel mechanism to enhance backdoor defense strategies within the deepfake detection framework.Several key factors emerged as significant in attacking and defending deepfake detection models.Firstly, the common practice of cropping faces in deepfake detection may result in triggers being either incomplete or absent in the cropped images.Secondly, deepfake detection is an inherently complex task, particularly when triggers are present, as it requires models to learn subtle features.Lastly, the number of classes in a classification task plays a crucial role in developing defenses against backdoor attacks.Based on these observations, we proposed enhancements to existing backdoor defense strategies, which demonstrated clear effectiveness in the deepfake detection context.This study advances our understanding of backdoor attacks within deepfake detection, emphasizing the need for more robust defense mechanisms.7.1 NAD defense result against BadNet attack in 6-class-case . . . . .7.2 NAD defense result against BadNet attack in binary case . . . . .7.3 BadNet attack without defense in binary case . . . . . . . . . . . .7.4 BadNet attack without defense in 6-class-case . . . . . . . . . . . .
Author: Qiu, Yuran
Language: English
Open Access status: green
OpenAlex ID: W7208690899
Cites: 36
Cited by:
Related to: 10