Vision-Language Models (VLMs), with their advantages in vision and language processing, exhibit immense potential in mobile intelligent systems. Integrating federated learning with parameter-efficient fine-tuning of VLMs helps address data heterogeneity challenges. However, existing methods mainly focus on task-specific patterns, neglecting the impact of general features, such as background information and low-quality data, which weakens the model's ability to generalize when handling data from different sources and dealing with fluctuations in quality. To tackle these challenges, we propose Inverse Feature Consistency Federated Unlearning (IFCFU) for VLM, comprising three components: 1) Feature Consistency Federated Learning (FCFL) aligns fine-tuned features with pre-trained features through constraints to ensure the preservation of general features; 2) Pseudo-label Low-quality Data Detection (PLDD) identifies potential low-quality data through model quality assessment and pseudo-label generation; 3) Inverse Feature Consistency Unlearning (IFCU) distances low-quality data features from optimal model features to eliminate the negative impact and restores training with pseudo-labels. Evaluations on StanfordCars show that FCFL increased accuracy by 4.97% and 29.88% under normal data and low-quality data configurations, respectively. PLDD identified over 90.00% of low-quality data, while IFCU improved the global model's accuracy by 4.43% with 80% low-quality data.

Federated unlearning (FUL) facilitates targeted unlearning of data-specific artifacts from trained federated learning (FL) models. Existing methodologies primarily emphasize individual client demands at a time, such as enforcing the “right to be forgotten” or mitigating poor-quality data contributions from a single client. These strategies frequently rely on client-side computational engagement during unlearning. However, these approaches often neglect situations in which it is crucial to unlearn the contributions of multiple participants at a time from a global model. Therefore, we introduce a novel FUL method named “FULClean” to parallelly erase the adverse impacts of poor-quality data from multiple participants without using client resources and maintaining global model performance. FULClean employs a novel Contrastive Threshold-based Contribution Rectification (CTCR) mechanism, that (1) identifies and classifies the local models into unaffected and potentially affected local models, (2) computes the dynamic contribution threshold for each potentially affected model based on layer-wise parameter divergence, and (3) selectively replaces or remove contributions based on their affected ratio. FULClean can categorize and swiftly eliminate the impacts of poor-quality data from multiple participants on the global model. This can be accomplished without communication and utilization of the client's resources. Extensive experiments are also conducted on four distinct datasets with two different models to showcase the effectiveness and efficiency of FULClean.

Federated unlearning (FUL) is a solution aimed at addressing the problem of removing data contributions from trained federated learning (FL) models. Existing FUL methods only focus on iterative unlearning of clients’ contributions and fail to perform unlearning in scenarios where multiple clients request to remove their data at a time. Additionally, FUL still needs to address issues, including convergence speed, maintaining the global model’s performance, and parallel unlearning to expedite the unlearning process. To fill this gap, we introduce Federated Clients Forgetting (FedCF), a fast and accurate FUL method that can eliminate single client contributions similar to existing methods, eliminate multiple clients’ contributions on the global model parallelly, ensure the performance of the unlearned global model, and reduce the unlearning time. The key idea is to construct a temporary model by extracting knowledge from the remaining clients’ updates and adding it to the corresponding parameters of the initial global model and then leverage a temporary model to reconstruct the unlearned global model. Extensive experiments on three benchmark datasets, FedCF demonstrates its efficiency and effectiveness for single client contribution unlearning, achieving an average time efficiency of 8.3x, 6.5x, and 4.1x over existing methods FedRetrain, FedEraser, and FUL with knowledge distillation, respectively. Additionally, FedCF showcases the time efficiency and performance guarantee after unlearning the contributions of multiple clients in parallel.

Due to its practical integration with IoT networks and ability to remove specific data contributions from a trained model, federated unlearning (FUL) has become increasingly attractive for IoT sensing systems. However, existing FUL methods are inadequate for rollback (unlearning) in IoT sensing systems due to IoT devices' heterogeneity and the training process's dynamic nature. To address these issues, we propose “FedRollback”, a lightweight novel FUL method for efficient rollback in IoT sensing systems. The key idea is to compress and filter the retaining updates, dynamically adjust the adjustment retraining rounds, and weighted aggregation based on each user (IoT device) resource availability score. This approach reduces computational overhead, maintains model performance, mini-mizes storage overhead, and speeds up the rollback process by retraining the model in a few rounds based on adaptive intervals. Evaluations on four distinct datasets (i.e., MNIST, KMNIST, SVHN, and CelebA) show an approximately 11x average speed-up over retraining from scratch and 5x and 4x improvement over FedEraser and FedRemover, demonstrating its efficiency and robustness in real-world IoT data sensing systems.

Federated learning (FL) offers a credible solution for distributed data trading since it could train machine learning models in a distributed manner thereby enhancing data privacy without sharing local data. However, it is still challenging to trade data through FL due to unfair model allocation issues arising from the unreliability of user-provided data. To tackle it, we propose F2UL, a Fairness-aware Federated UnLearning solution that distributes models to users commensurate with their data quality. F2UL is trained in a two-stage (TST) way and contains three main components: 1) Label-free model quality assessment (LMQA) promotes fairness by evaluating models without user-specific data, ensuring uniform assessment standards. 2) Fair model distribution (FMD) addresses the issue of unfair model distribution by allocating models with feature mapping deviation, ensuring that users who contribute low-quality models do not receive enhanced models. 3) User data federated unlearning (UDFU) ensures fairness in model distribution by employing rapid recovery federated unlearning, safeguarding regular users from the adverse effects of low-quality data on model performance. In experiments on CIFAR10, F2UL reduces the low-quality data user accuracy to 9.09% and increases regular users’ accuracy by 3.02%, thereby demonstrating F2UL's capacity to ensure fairness in data trading.