Abstract:Federated Class Incremental Learning (FCIL) is a critical yet largely underexplored issue that deals with the dynamic incorporation of new classes within federated learning (FL). Existing methods often employ generative adversarial networks (GANs) to produce synthetic images to address privacy concerns in FL. However, GANs exhibit inherent instability and high sensitivity, compromising the effectiveness of these methods. In this paper, we introduce a novel data-free federated class incremental learning framework with diffusion-based generative memory (DFedDGM) to mitigate catastrophic forgetting by generating stable, high-quality images through diffusion models. We design a new balanced sampler to help train the diffusion models to alleviate the common non-IID problem in FL, and introduce an entropy-based sample filtering technique from an information theory perspective to enhance the quality of generative samples. Finally, we integrate knowledge distillation with a feature-based regularization term for better knowledge transfer. Our framework does not incur additional communication costs compared to the baseline FedAvg method. Extensive experiments across multiple datasets demonstrate that our method significantly outperforms existing baselines, e.g., over a 4% improvement in average accuracy on the Tiny-ImageNet dataset.

What problem does this paper attempt to address?

### Problems the Paper Aims to Solve This paper aims to address the issue of catastrophic forgetting in Federated Class Incremental Learning (FCIL). Specifically, FCIL is a critical problem in the Federated Learning (FL) environment where new classes are dynamically incorporated. Existing methods typically use Generative Adversarial Networks (GANs) to generate synthetic images to address privacy issues in FL. However, GANs have inherent instability, which affects the effectiveness of these methods. To overcome these issues, the authors propose a novel data-free federated class incremental learning framework called Data-Free Federated Class Incremental Learning with Diffusion-Based Generative Memory (DFedDGM). This framework generates stable, high-quality images through a diffusion model, thereby alleviating the problem of catastrophic forgetting. Additionally, the authors design a new balanced sampler to train the diffusion model to address the common Non-Independent and Identically Distributed (Non-IID) problem in FL and introduce an entropy-based sample filtering technique to improve the quality of generated samples. Finally, the authors combine knowledge distillation with feature-based regularization terms for better knowledge transfer. ### Main Contributions 1. **Proposed a novel data-free federated class incremental learning framework DFedDGM**: This framework alleviates the problem of catastrophic forgetting in FCIL through diffusion-based generative memory without requiring additional communication costs. 2. **Designed a new balanced sampler**: Helps train the diffusion model to address the prevalent Non-IID problem in FL. 3. **Introduced an entropy-based sample filtering method**: Removes low-quality generated samples from an information theory perspective. 4. **Conducted extensive experiments**: Validated the effectiveness of the method on multiple datasets, showing that it significantly outperforms existing FCIL methods, such as improving the average accuracy on the Tiny-ImageNet dataset by over 4%. ### Related Work - **Continual Learning (CL)**: CL studies how to address the problem of catastrophic forgetting in machine learning. In the CL framework, training data is presented to the model in the form of a sequence of tasks, with the model only having access to one dataset (task) at each time step and needing to perform well on both the current task and previous tasks. - **Federated Learning (FL)**: FL is a distributed learning framework that builds a global model on a central server by aggregating parameters independently learned on multiple client devices. The performance of FL is limited by the dispersion of data (Non-IID). - **Federated Continual Learning (FCL)**: FCL addresses the problem of catastrophic forgetting in the FL environment, especially when data classes are constantly changing. Existing FCL methods include using generative models to generate synthetic data to reduce forgetting. ### Problem Definition - **Federated Learning (FL)**: Assume there are \( N \) different clients, each client \( i \) has its own private dataset \( D_i \). The goal of FL is to collaboratively develop a global model \( \theta_g \) without sharing private datasets. - **Class Incremental Learning (Class-IL)**: Consider a set of tasks \( T \), each task \( T_t \) contains a dataset \( D_t \). Each new task \( T_t \) introduces new classes, and the goal is to develop a model \( \theta_T \) that performs well not only on the last task \( T_T \) but also maintains performance on all previous tasks, avoiding catastrophic forgetting.

Data-Free Federated Class Incremental Learning with Diffusion-Based Generative Memory

Overcoming Catastrophic Forgetting in Federated Class-Incremental Learning via Federated Global Twin Generator

Federated Class-Incremental Learning with New-Class Augmented Self-Distillation

Privacy-Preserving Federated Class-Incremental Learning

TARGET: Federated Class-Continual Learning Via Exemplar-Free Distillation

FedSKF: Selective Knowledge Fusion via Optimal Transport in Federated Class Incremental Learning

Diffusion-Driven Data Replay: A Novel Approach to Combat Forgetting in Federated Class Continual Learning

Don't Memorize; Mimic The Past: Federated Class Incremental Learning Without Episodic Memory

Federated Class-Incremental Learning with Hierarchical Generative Prototypes

FedINC: an Exemplar-Free Continual Federated Learning Framework with Small Labeled Data

FL-Clip: Bridging Plasticity and Stability in Pre-Trained Federated Class-Incremental Learning Models

Federated Class-Incremental Learning with Prompting

Few-Shot Class-Incremental Learning Via Class-Aware Bilateral Distillation

Using Diffusion Models as Generative Replay in Continual Federated Learning -- What will Happen?

Communication-Efficient Federated Data Augmentation on Non-IID Data

RRA-FFSCIL: Inter-intra classes representation and relationship augmentation federated few-shot incremental learning

A Cognition-Driven Framework for Few-Shot Class-Incremental Learning

ICFD: an Incremental Learning Method Based on Data Feature Distribution.

Text-Enhanced Data-free Approach for Federated Class-Incremental Learning

Federated Generative Learning with Foundation Models

FSCIL-EACA: Few-Shot Class-Incremental Learning Network Based on Embedding Augmentation and Classifier Adaptation for Image Classification