악의적 클라이언트의 모델 공격에 강건한 확률적 연합 언러닝 기법

다국어 초록 (Multilingual Abstract)

Federated learning (FL) is a distributed learning paradigm that enables the training of a global model without directly sharing clients’local data, thereby preserving data privacy. However, in federated learning environments, malicious clients can launch model poisoning attacks such as backdoor attacks, which severely compromise the integrity and reliability of the global model. To mitigate these threats, federated unlearning (FU) has been proposed; nevertheless, most existing approaches rely on deterministic binary decisions that classify clients as either benign or malicious. Such deterministic decisions are vulnerable to detection uncertainty, leading to false positives or false negatives, which in turn cause unnecessary removal of benign client contributions or residual influence of malicious updates.
To overcome these limitations, this thesis proposes a probabilistic federated unlearning approach that interprets malicious-client detection outcomes as probabilities and directly incorporates them into the unlearning process. The proposed method transforms each client’s maliciousness probability through a hyperbolic tangent (tanh)–based nonlinear function to define weights that simultaneously represent benign/malicious tendency and contribution strength. Based on these probabilistic weights, we design unlearning schemes based on probabilistically weighted gradient descent, gradient ascent, and a combined gradient descent–ascent strategy.
Extensive experiments conducted on the CIFAR-10 dataset under backdoor attack scenarios demonstrate that the proposed probabilistic federated unlearning methods maintain global model accuracy comparable to deterministic baselines while achieving a more stable and effective reduction in attack success rate under detection uncertainty. Furthermore, by adjusting the tanh scaling parameter, the behavior of the proposed framework gradually converges to that of deterministic unlearning, indicating that the proposed approach constitutes a generalized probabilistic federated unlearning framework that subsumes existing deterministic methods. These results suggest that the proposed framework provides a more robust and stable solution for federated unlearning in realistic federated learning environments.

목차 (Table of Contents)

• Ⅰ. 서론 1
• Ⅱ. 관련연구 4
• 1. 연합학습 4
• 2. 연합 언러닝 5
• 3. 머신 언러닝 알고리즘 6
• 4. 연합 언러닝 알고리즘 7
• Ⅲ. 본론 11
• 1. 연구개요 11
• 2. 연구방법 11
• Ⅳ. 실험및결과 19
• 1. 실험데이터 19
• 2. 백도어 공격 시나리오 19
• 3. 실험설계 및 평가지표 20
• 4. 실험 결과 23
• Ⅴ. 결론 46
• * 참고문헌 47
• ABSTRACT 50