Federated Class Incremental Learning: A Pseudo Feature Based Approach Without Exemplars

Min Kyoon Yoo, Yu Rang ParkarXiv 2026 · DOI · arXiv

TL;DR

The paper introduces FCLPF (Federated Class Incremental Learning with Pseudo Features), a novel framework that allows decentralized models to learn new classes over time without forgetting old ones. By generating "pseudo features" using Principal Component Analysis (PCA) on class prototypes, FCLPF eliminates the need to store privacy-sensitive old data or train computationally expensive generative models like GANs.

What Problem Does This Paper Address?

Standard Federated Learning (FL) assumes that the data across decentralized clients is static. However, in the real world—like in healthcare—new types of data (e.g., new diseases) arrive continuously. When a model is updated with this new information, it often suffers from catastrophic forgetting, where the weights shift so much to accommodate new classes that performance on previously learned classes plummets.

Existing solutions for "Class Incremental Learning" (CIL) usually rely on rehearsal memory, where a small subset of old data (exemplars) is stored. In a federated setting, this is a non-starter: it consumes precious local storage and raises significant privacy concerns. While some researchers have tried using Generative Adversarial Networks (GANs) to "imagine" old data, these models are notoriously difficult to train on edge devices and require massive communication bandwidth to sync between clients and servers.

Methods

FCLPF moves away from raw data and heavy generative models, focusing instead on Prototypes—the average feature vectors of a class.

The core of the method involves four steps:

  1. Prototype Aggregation: Clients send class-specific prototypes to the server, which averages them to create global prototypes.
  2. Closest Prototype Matching: To "reconstruct" an old class, the system identifies the most similar new class using cosine similarity.
  3. PCA Projection: The authors apply PCA to the distribution of old and new prototypes. By projecting the difference between an old prototype and its closest new neighbor onto these principal axes, they capture the essential "translation" needed to turn a new feature into a pseudo-old feature.
  4. Pseudo Feature Generation: During local training, clients take features from current data and apply the PCA-derived translation to create synthetic "pseudo features" of old classes. This allows the model to "rehearse" old knowledge without ever seeing the original data.

To keep things efficient, FCLPF freezes the feature extractor (the backbone of the model) after the initial task. Only the final Fully Connected (FC) layer and the small prototype vectors are exchanged between the client and server.

Fig. 1: Overview of our method. ( P_n ) is the prototype of the nth class, and FC de- notes the fully connected (FC) layer. F(x_n) represents the feature of x_n, whereas G(x_p) denotes the pseudo feature for the past class. Only prototypes and FC layers are ex- changed between the client and server. The server aggregates the prototype and FC layer.

Datasets

The authors evaluated FCLPF across three major benchmarks, using various "incremental scenarios" (e.g., starting with 50 classes and adding 10 classes at a time):

  • CIFAR-100: A standard image classification dataset with 100 classes.
  • TinyImageNet: A subset of ImageNet with 200 classes at a lower resolution.
  • SuperImageNet: Used to further demonstrate the model's robustness in complex scenarios.

Key Insights and Ideas

  • Feature Translation over Data Generation: The "aha moment" here is that you don't need a GAN to recreate an image if you can mathematically translate the features of a similar new class into the feature space of an old one.
  • PCA as a Privacy Shield: By using PCA on prototypes, the method extracts the "essence" of class differences without needing raw samples, making it highly compatible with the privacy-first nature of federated learning.
  • Efficiency through Freezing: Freezing the backbone model significantly slashes communication costs. As shown in the paper's analysis, FCLPF exchanges orders of magnitude fewer parameters than GAN-based methods like MFCL.

Key Results

FCLPF consistently outperformed baselines like FedAvg, FedProx, and the state-of-the-art MFCL across all metrics:

  • Accuracy: On CIFAR-100, FCLPF achieved an average accuracy of 51.87%, beating the next best baseline (MFCL) at 47.72%.
  • Forgetting: The "Average Forgetting" metric—which measures how much accuracy is lost on old tasks—was remarkably low. On TinyImageNet, FCLPF showed only 3.14% forgetting, compared to a staggering 24.46% for the baseline.
  • Scalability: As the number of incremental tasks increased, the performance gap between FCLPF and other methods widened, suggesting it is much more stable for long-term learning.
Fig. 2: Comparison of model performance across different tasks and datasets

Limitations and Future Work

While FCLPF is highly efficient, the authors acknowledge a few areas for improvement:

  • Frozen Backbone: Freezing the feature extractor after the first task assumes the initial data is diverse enough to learn a general feature representation. If later classes are radically different, the model might struggle.
  • Prototype Averaging: Currently, prototypes are simple averages. The authors suggest exploring more sophisticated ways to aggregate and apply prototypes to improve the robustness of pseudo features.
  • Future Directions: The team plans to investigate ways to fine-tune the feature extractor alongside pseudo features to further boost performance without re-introducing catastrophic forgetting.

Summary updated

Abstract

Federated learning often assumes that data is fixed in advance, which is unrealistic in many real-world scenarios where new data continuously arrives, causing catastrophic forgetting. To address this challenge, we propose FCLPF (Federated Class Incremental Learning with Pseudo Features), a method that uses pseudo features generated from prototypes to mitigate catastrophic forgetting. Our approach reduces communication costs and improves efficiency by eliminating the need for past data and avoiding computationally heavy models like GANs. Experimental results on CIFAR-100 show that FCLPF achieves an average accuracy of 51.87% and an average forgetting of 9.62%, significantly outperforming existing baselines with an average accuracy of 47.72% and forgetting of 20.46%. On TinyImageNet, FCLPF achieves 37.56% accuracy and 3.14% forgetting, compared to the baselines' 27.69% accuracy and 24.46% forgetting, demonstrating the superior performance of FCLPF.
BibTeX
@misc{yoo2026federated,
  author = {Min Kyoon Yoo and Yu Rang Park},
  title = {{Federated Class Incremental Learning: A Pseudo Feature Based Approach Without Exemplars}},
  year = {2026},
  eprint = {2611.15025},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15025},
}
Figure