💻 computer science

EPACE-FL: Efficient Privacy-Preserving And Communication in Edge Computing of Federated Learning

The EPACE-FL framework addresses the challenges of privacy leakage and low communication efficiency in edge-based federated learning by introducing a Dynamic Quantization and Chaotic Masking (DQCM) algorithm that jointly compresses and encrypts model parameters, significantly reducing communication overhead while enhancing protection against malicious attacks.

Original authors: Chao-Hsien Hsieh, Peidong Wang

Published 2026-10-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chao-Hsien Hsieh, Peidong Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital age, our lives are increasingly recorded as streams of data: the heart rate from a smartwatch, the route of a self-driving car, or the medical history stored on a hospital server. For decades, the standard way to turn this raw information into useful knowledge was to gather it all in one place, a massive cloud server where powerful computers could analyze it. However, this centralization creates a dangerous bottleneck. Sending vast amounts of personal data across the internet is slow, expensive, and risky, as every transmission offers a chance for private details to be stolen or misused. A newer approach called federated learning has emerged to solve this by keeping the data where it lives. Instead of sending the data to a computer, the computer is sent to the data. The learning happens locally on the device, and only a small, summarized update of what was learned is sent back to the central server. This keeps the raw information safe, but it introduces its own set of problems. The devices at the edge of the network, like sensors or phones, often have weak batteries and unstable internet connections. Sending even these small updates can clog the network, and the updates themselves can sometimes be reverse-engineered by hackers to reveal the private data they were meant to protect.

Researchers Chao-Hsien Hsieh and Peidong Wang have developed a new system called EPACE-FL to navigate these difficult conditions. Their work focuses on the specific challenges of edge computing, where devices operate with limited power and fluctuating network speeds. The team identified that existing methods for protecting privacy, such as complex mathematical encryption, are too heavy for these small devices, while simple compression techniques often fail to stop hackers from reconstructing private information. To bridge this gap, they created a framework that combines two distinct actions into a single, streamlined process. First, the system takes the large, complex numbers that represent a device's learning and shrinks them down into much smaller, simpler integers. This is not a static reduction; the system constantly monitors the internet connection. If the network is strong, it keeps the numbers precise. If the connection is weak or the bandwidth drops, it immediately compresses the data further to ensure the message gets through without timing out.

Once the data is compressed, the system applies a second layer of protection that is unique to this approach. Instead of using heavy, slow encryption methods, the device generates a random, chaotic pattern to mix with the compressed numbers. This pattern acts like a one-time pad, scrambling the information so that even if a hacker intercepts the transmission, they see only meaningless noise. Crucially, this scrambling is designed so that when all the devices send their scrambled updates to the central server, the random patterns cancel each other out perfectly. The server can then add up the true learning from all the devices to improve the global model, but it never sees the individual, unscrambled data from any single user. This method allows the system to operate efficiently on devices with limited power, such as those simulated in the study using Raspberry Pi microcomputers, while maintaining a high level of security.

The researchers tested their system using standard image-recognition datasets, training models to identify handwritten digits and everyday objects. They compared their method against existing techniques, including those that rely on adding random noise to protect privacy or those that use heavy encryption. The results showed that the new framework significantly reduced the amount of data that needed to be sent over the network. Specifically, the total communication overhead for each client was reduced by 29.3 percent compared to standard methods. This efficiency was achieved without sacrificing the accuracy of the final model, which remained nearly as good as if the data had been sent in its original, unencrypted form. Furthermore, when the system was pitted against sophisticated attacks designed to reverse-engineer private data from the transmitted updates, it proved highly effective. The framework successfully blocked these attacks 41.6 percent more often than a leading privacy-preserving baseline, effectively preventing the reconstruction of sensitive information.

The study demonstrates that it is possible to have both speed and security in distributed learning, provided the methods are tailored to the constraints of the real world. By dynamically adjusting the size of the data packets to match the available internet speed and using lightweight, chaotic scrambling to hide the content, the EPACE-FL framework offers a practical solution for the future of edge intelligence. It suggests that we do not need to choose between protecting our privacy and training powerful artificial intelligence models; with the right design, we can do both simultaneously, even on devices with limited resources and unstable connections. The work lays a foundation for more robust and efficient learning systems that can operate securely in the complex, variable environments of the modern digital landscape.

Technical Summary of EPACE-FL: Efficient Privacy-Preserving And Communication in Edge Computing of Federated Learning

Problem Statement
The paper addresses two critical bottlenecks in deploying Federated Learning (FL) within edge computing environments: insufficient privacy protection and low communication efficiency. While FL keeps raw data local, conventional approaches still transmit model parameters (often in 32-bit floating-point format) over unstable wireless channels. This creates two distinct issues:

Communication Constraints: Edge devices face limited battery life, heterogeneous computational resources, and highly volatile network bandwidth. Transmitting high-dimensional model parameters frequently leads to network congestion, intolerable latency, and training paralysis.
Privacy Vulnerabilities: Although raw data is not shared, transmitted model parameters (gradients/updates) can be intercepted by malicious clients or eavesdroppers. Attackers can utilize gradient inversion attacks (e.g., EGIA) to reconstruct sensitive private information from these parameters.

Existing solutions often fail to address these issues simultaneously. Techniques like Differential Privacy (DP) degrade model accuracy, while Homomorphic Encryption (HE) and Secure Multi-Party Computation (SMC) impose computational and communication overheads too heavy for resource-constrained edge devices. Furthermore, many existing methods treat model compression and privacy protection as separate, sequential steps, failing to achieve synergistic optimization.

Methodology: The EPACE-FL Framework
To resolve these limitations, the authors propose EPACE-FL, a framework specifically tailored for edge computing that integrates Dynamic Quantization and Chaotic Masking (DQCM). The framework operates on a standard FL paradigm (one server, multiple clients) but modifies the parameter transmission process through three core mechanisms:

Dynamic Quantization (Communication Efficiency):
Instead of fixed-bit quantization, the framework dynamically adjusts the quantization bit-width ( $b_t$ ) based on real-time network bandwidth and communication budgets.
- Floating-point model updates are clipped to a shared range $[-R_t, R_t]$ and quantized into low-bit integers.
- When bandwidth is low, the system selects a lower bit-width (e.g., 4 bits) to minimize payload size. When bandwidth is high, it increases the bit-width (e.g., 8 bits) to preserve model accuracy.
- This reduces the transmission payload ( $\mu$ ) significantly without the ciphertext expansion associated with encryption.
Chaotic Masking (Privacy Protection):
To prevent privacy inference attacks on the quantized integers, the framework employs a lightweight additive masking scheme based on chaotic pseudo-random sequences.
- Pairwise Key Generation: Clients establish pairwise shared secret keys ( $K_{i,j}$ ) to generate domain-separated seeds for model updates and aggregation weights.
- Mask Generation: Using a Chaotic Masking Generator (CMG), clients generate random integer masks. These masks are added to the quantized parameters modulo a large prime $p$ .
- Secure Aggregation: Due to the properties of the chaotic generator and the specific construction of total masks (where $\sum r_i = 0$ ), the masks cancel out during server-side aggregation. The server obtains the exact sum of the parameters without ever seeing individual plaintext updates.
Distance-Aware Aggregation:
To mitigate the accuracy loss caused by quantization and statistical heterogeneity (Non-IID data), the framework introduces a distance-aware weighting mechanism.
- Clients calculate the Euclidean distance between their local model and the current global model.
- An adjustment coefficient is applied to the aggregation weights based on this distance. Models that deviate significantly from the global average receive lower weights, reducing the impact of quantization errors and statistical outliers on the global model.

Key Contributions
The paper claims three primary contributions:

EPACE-FL Framework: A novel architecture designed specifically for edge constraints that simultaneously minimizes computational and communication overheads while ensuring privacy.
DQCM Algorithm: A joint compression and encryption algorithm that dynamically adapts to network conditions and uses chaotic masking to secure quantized parameters, avoiding the high overhead of traditional cryptographic methods.
Comprehensive Evaluation: A rigorous experimental validation demonstrating that the framework achieves high-accuracy model aggregation with significantly reduced communication costs and enhanced privacy defense.

Experimental Results
The authors evaluated EPACE-FL on MNIST (LeNet-5) and CIFAR-10 (ResNet-18) datasets under both IID and Non-IID settings, simulating edge conditions using Raspberry Pi 4 clients and an NVIDIA RTX 4090 server. Key findings include:

Communication Efficiency: Compared to the baseline FedAvg, EPACE-FL reduced the total communication overhead per client by 29.3% (from 177.08 GB to 119.11 GB cumulative traffic).
Privacy Defense: Against Gradient Inversion Attacks (EGIA), EPACE-FL achieved a defense success rate of 89.4%, representing a 41.6% improvement over the DP-FedAvg baseline. It effectively blocked the reverse reconstruction of private information.
Model Utility: The framework maintained convergence rates and final accuracy comparable to plaintext FedAvg, significantly outperforming DP-FedAvg and static quantization methods (AGQ-FL).
Computational Overhead: The execution time on edge devices was marginal compared to FedAvg and significantly lower than Fully Homomorphic Encryption (FHE) approaches.

Significance and Claims
The paper asserts that EPACE-FL fundamentally resolves the trade-off between privacy and efficiency in edge computing. By synergizing dynamic quantization and chaotic masking, the framework avoids the "plaintext exposure" risk of compression-only methods and the "computational explosion" of encryption-only methods. The authors claim this approach enables the deployment of robust, privacy-preserving FL in real-world edge scenarios characterized by unstable networks and limited hardware resources, such as healthcare monitoring and industrial IoT, without sacrificing model utility. Future work is noted to focus on resilience against active client dropouts and poisoning attacks.

Technical Summary of EPACE-FL: Efficient Privacy-Preserving And Communication in Edge Computing of Federated Learning

More like this