Large Language Models (LLMs) increasingly leverage Federated Learning (FL) to utilize private, task-specific datasets for fine-tuning while preserving data privacy. However, while federated LLM frameworks effectively enable collaborative training without raw data sharing, they critically lack built-in mechanisms for regulatory compliance like GDPR’s right to be forgotten. Integrating private data heightens concerns over data quality and long-term governance, yet existing distributed training frameworks offer no principled way to selectively remove specific client contributions post-training. Due to distributed data silos, stringent privacy constraints, and the intricacies of interdependent model aggregation, federated LLM unlearning is significantly more complex than centralized LLM unlearning. To address this gap, we introduce Oblivionis, a lightweight learning and unlearning framework that enables clients to selectively remove specific private data during federated LLM training, enhancing trustworthiness and regulatory compliance. By unifying FL and unlearning as a dual optimization objective, we incorporate 6 FL and 5 unlearning algorithms for comprehensive evaluation and comparative analysis, establishing a robust pipeline for federated LLM unlearning. Extensive experiments demonstrate that Oblivionis outperforms local training, achieving a robust balance between forgetting efficacy and model utility, with cross-algorithm comparisons providing clear directions for future LLM development.

With the widespread deployment of Low Earth Orbit (LEO) satellites, they generate a vast amount of data. This data has been instrumental in supporting machine learning (ML) in various terrestrial services to address global challenges such as monitoring climate change and natural disasters. However, many national regulations restrict the direct transmission of satellite data to ground stations (GSs). Therefore, ground-assisted satellite federated learning (FL) has emerged as a paradigm to safeguard data privacy by transferring model parameters instead of raw data for collaborative training. At present, the existing groundassisted satellite FL methods encounter practical challenges: 1) The dynamic environment of LEO satellites results in continuous changes in the types of data collected by satellites, making it difficult for traditional FL models to adapt to these changes. This can lead to a deterioration in model accuracy over extended periods of model training. 2) Communication between satellites and GS is affected by atmospheric interference and weather factors, resulting in increased transmission delays and affecting the realtime efficiency of the FL system. In response to these challenges, we propose a dynamic, efficient, and distributed ground-assisted LEO satellite federated learning (DEDFL) framework to improve model accuracy and reduce satellite communication delays. In DEDFL, we design a Balanced Class Memory Extraction and an information playback strategy that enables the onboard FL model to adapt to changing satellite data types, thus achieving a performance balance across different classes. Additionally, we propose an adaptive fine coding method for parameter adoption prior to satellite transmission, effectively reducing the delay caused by satellites and ground-specific environmental variations. Experimental results demonstrate that the DEDFL method offers better accuracy and communication efficiency than other baseline algorithms.

Federated Learning (FL), as a distributed machine learning approach, makes it possible to realize global training and knowledge sharing of models while protecting data privacy. However, existing FL methods still face some challenges. First, since the local client model has only a limited parameter capacity when the model learns a new task or new data, previously learned information can be lost quickly and in large amounts, leading to catastrophic forgetting. Second, the private information may still be leaked by analyzing the parameters uploaded by the clients. Besides, the data distributions of FL clients are usually non-independent and identically distributed, leading to a degradation of the quality of the global aggregation model. These challenges pose obstacles to the effective implementation of federated learning. To address the above issues, we propose a novel differentially private federated class-incremental learning (DPFCIL) method. Specifically, it includes three aspects: 1) We design an elastic memory strategy that mitigates the class incremental model’s forgetting of old knowledge by dynamically expanding new modules and applying network regularization at the local client. 2) Before uploading parameters to the local client, we introduce differential privacy to prevent information leakage effectively. 3) We propose to use a dynamic adaptive aggregation method that relies on model parameter differences in the server aggregation phase to mitigate the loss of global accuracy due to the heterogeneity of the client’s data. Extensive experiments are designed to compare with state-of-the-art baseline methods on several datasets, and the experimental results show that DPFCIL outperforms other methods by an average of at least 5.6% in accuracy.

In the past few years, numerous Deep Neural Network (DNN) models and frameworks have been developed to tackle the problem of real-time object detection from RGB images. Ordinary object detection approaches process information from the images only, and they are oblivious to the camera pose with regard to the environment and the scale of the environment. On the other hand, mobile Augmented Reality (AR) frameworks can continuously track a camera's pose within the scene and can estimate the correct scale of the environment by using Visual-Inertial Odometry (VIO). In this paper, we propose a novel approach that combines the geometric information from VIO with semantic information from object detectors to improve the performance of object detection on mobile devices. Our approach includes three components: (1) an image orientation correction method, (2) a scale-based filtering approach, and (3) an online semantic map. Each component takes advantage of the different characteristics of the VIO-based AR framework. We implemented the AR-enhanced features using ARCore and the SSD Mobilenet model on Android phones. To validate our approach, we manually labeled objects in image sequences taken from 12 room-scale AR sessions. The results show that our approach can improve on the accuracy of generic object detectors by 12% on our dataset.