Large language models are increasingly deployed across mobile and edge environments, where privacy-sensitive and heterogeneous user data raise critical concerns of copyright infringement, data leakage, and regulatory non-compliance. Ma chine unlearning has thus emerged as an essential capability to remove the influence of specific data without full retraining. However, two key challenges remain open: 1) how to quantify unlearning to enable data valuation without retraining, especially since the massive scale of pretraining makes it infeasible to evaluate the contribution of individual data samples in advance, and 2) how to verify the correctness without retraining to ensure that third-party auditors can efficiently confirm the complete removal of targeted data influence. To address the aforementioned challenges, in this paper, we design a dual-stage machine unlearning framework to quantify the contribution of forgotten data and certify data removal without full retraining, serving as an auditing layer for first-order unlearning methods. Specifically, we design a run-time Shapley value-based unlearned data evaluation mechanism that utilizes a first-order approximation strategy to estimate the marginal contribution of forgotten samples. Moreover, we propose a proof of unlearning mechanism that generates compact, auditable artifacts of the unlearning process to efficiently verify that the targeted data influence has been completely removed. Compared with five state-of-the-art unlearning baselines, our approach achieves effectiveness in data valuation, stronger guarantees of removal correctness, and lower computational overhead.

Federated Learning has garnered significant attention in practical applications due to its privacy-preserving properties but faces serious threats from backdoor attacks. Current defenses primarily rely on server-side anomaly detection and robust aggregation, but lack systematic strategies for proactively erasing backdoors from the perspective of attackers. To bridge this gap, we propose an efficient and stealthy federated backdoor removal framework. Specifically, our method incorporates a two-stage training approach: reinforced negative learning and positive memory recovery. In addition, we intro-duce a hybrid regularization strategy that combines dynamic L1 regularization with Elastic Weight Consolidation, together with a synchronized differential amplification mechanism for both weights and buffers and a global norm clipping strategy. These components collectively effectively erase backdoor effects, maintain main-task accuracy, and significantly reduce detection risk on the server side. Experimental evaluations demonstrate that, compared to existing backdoor unlearning methods, our approach decreases the success rate of the backdoor attack by up to 11% without compromising the primary precision. Furthermore, our method substantially improves the stealthiness of the update, reducing the L2 norm fluctuations to less than 33% of the baseline levels.

Dataset distillation is utilized to condense large datasets into smaller synthetic counterparts, effectively reducing their size while preserving their crucial characteristics. In Federated Learning (FL) scenarios, where individual devices or servers often lack substantial computational power or storage capacity, the use of dataset distillation becomes particularly advantageous for processing large volumes of data efficiently. Current research in dataset distillation for FL has primarily focused on enhancing accuracy and reducing communication complexity, but it has largely neglected the potential risk of backdoor attacks. To solve this issue, in this paper, we propose three adaptive dataset condensation based backdoor attacks against dataset distillation for FL. Adaptive attacks in dataset distillation for FL dynamically modify triggers during the training process. These triggers, embedded in the synthetic data, are designed to bypass traditional security detection. Moreover, these attacks employ self-adaptive perturbations to effectively respond to variations in the model's parameters. Experimental results show that the proposed adaptive attacks achieve at least 5.87% higher success rates, while maintaining almost the same clean test accuracy, compared to three benchmark methods.

With the advancement of federated crowdsourcing services, the associated privacy concerns have attracted growing attention from both academia and industry. Existing privacy laws impose strict requirements concerning the right to be forgotten for data used in training AI models. In federated crowdsourcing services, the right to be forgotten is guaranteed through federated unlearning. Current federated unlearning solutions encompass a two-step process: first, eliminating model updates associated with the target data to achieve unlearning, followed by retraining among the remaining clients to restore the performance of federated crowdsourcing services. However, this indiscriminate removal of model updates, while safeguarding the privacy of the target data, also greatly undermines the generalization performance of the global model. Moreover, relying on client-side retraining imposes additional economic costs on the federated crowdsourcing service. To tackle the above issues, this paper proposes an efficient federated unlearning framework for federated crowdsourcing services, which is based on adaptive parameter clipping and data-free distillation. We first compute the Fisher information matrix (FIM) to approximate the correlation between the target data and all model parameters, which is utilized to adaptively clip each parameter of the global model. Then, we model the softmax layer of the global model to synthesize pseudo-samples, enabling the retrain process on the crowdsourcing platform for the recovery of generalization performance. We conducted extensive experiments on three datasets, and the results demonstrate that our proposed framework not only possesses outstanding data removal capability but also outperforms the comparison methods in terms of computation time and storage space.