Research map / Federated learning and privacy / Personalized and heterogeneous FL

Federated Class-Incremental Learning: A Hybrid Approach Using Latent Exemplars and Data-Free Techniques to Address Local and Global Forgetting

ICLR 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated prompt and LoRA tuning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Hybrid Replay (HR), a customized VAE-style autoencoder that replays compact latent exemplars for local forgetting and decoder-generated synthetic data for global forgetting in federated class-incremental learning, with a Lennard-Jones potential to align class centroids.

Problem

Federated class-incremental learning suffers from local forgetting (class imbalance within each client, intra-client task confusion) and global forgetting (inter-client task confusion), while clients have tight memory and privacy constraints.

Approach

A single customized autoencoder serves as classifier and generator: clients store low-dimensional latent exemplars decoded back to input space for local replay, and the decoder synthesizes data from perturbed server-aligned class centroids (placed via Lennard-Jones potential) for data-free global replay.

Key results

• HR reaches 65.84% on CIFAR-100 (10/10/50/5) versus 61.83% for the best exemplar-based baseline GLFC and 63.47% for hybrid REMIND+ in the reported table.
• HR-mini with 10x less memory reaches 60.42% on CIFAR-100, comparable to exemplar-based baselines.
• HR without latent exemplars (data-free variant) reaches 52.27% on CIFAR-100, outperforming data-free baselines such as MFCL (50.03%).

Measured effects reported in the paper

Surprising findings

• Perfect (raw) exemplars perform only marginally better than latent exemplars, suggesting image quality for replay is not a major factor for overcoming forgetting.
• Removing global replay causes a severe drop (e.g., CIFAR-100 from 65.84% to 51.76%), indicating global forgetting dominates.

Limitations

• Decoder capacity trades off between replay realism and privacy; an overly complex decoder may memorize training samples.
• Larger decoders increase training computation, though not inference cost.
• Quantization is discussed as future work rather than implemented.

Experimental setup

Datasets: CIFAR-100 ImageNet-Subset TinyImageNet

Benchmarks: CIFAR-100 ImageNet-Subset TinyImageNet

Metrics: accuracy forgetting memory footprint wall-clock time

Baselines: iCaRL BiC PODNet DDE+iCaRL GeoDL+iCaRL SS-IL GLFC FedFTG FedSpace TARGET MFCL REMIND

Builds on: Remind VAE GLFC iCaRL

Closest papers by meaning

• Class-wise Balancing Data Replay for Federated Class-Incremental Learning (NeurIPS 2025 · similarity 0.90)
• Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.89)
• Federated Orthogonal Training: Mitigating Global Catastrophic Forgetting in Continual Federated Learning (ICLR 2024 · similarity 0.89)
• C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning (NeurIPS 2025 · similarity 0.88)
• Towards Efficient Replay in Federated Incremental Learning (CVPR 2024 · similarity 0.88)
• Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.88)