Skip to content

QFedAgent: Quantum-Enhanced Personalized Federated Learning for Multi-Agent Activity Recognition

🕒 Published (v1): 2026-07-02 16:54 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

QFedAgent proposes a hybrid quantum-classical personalized federated learning framework for multi-agent human activity recognition from IMU data. A variational quantum circuit (VQC) replaces classical MLP-based multimodal fusion, encoding accelerometer–gyroscope interactions via quantum entanglement with only 72 rotation parameters versus 33K in the MLP baseline. On the OPPORTUNITY dataset under non-IID subject partitions, the system achieves 97.7% mean accuracy with ~10× total fusion parameter reduction.

Problem

Standard federated learning (e.g., FedAvg) assumes approximately IID client data, but distributed robotic/wearable agents operate under heterogeneous motion patterns and sensor conditions, producing non-IID multimodal streams. Classical cross-modal fusion modules (attention, transformers, MLPs) introduce substantial parameter overhead that amplifies communication cost and complicates aggregation in heterogeneous federated settings.

Method

Each client runs dual CNN encoders that produce \(d\)-dimensional embeddings \(e_a, e_g \in \mathbb{R}^d\) from accelerometer and gyroscope channels respectively. The embeddings are projected to qubit rotation angles \(\phi = \pi \cdot \tanh(W_{\text{pre}}[e_a \| e_g]) \in [-\pi, \pi]^N\), then encoded into an \(N=8\) qubit circuit via AngleEmbedding (RX rotations). Cross-modal correlations are established through \(L=3\) StronglyEntanglingLayers parameterized by \(\theta^{\text{vqc}} \in \mathbb{R}^{L \times N \times 3}\). The fused representation \(z_i = \langle \psi(\phi,\theta) | Z_i | \psi(\phi,\theta) \rangle\) is mapped back to embedding space via \(W_{\text{post}}\) and layer normalization. Parameters are split into global (\(w_{\text{enc}}, \theta^{\text{vqc}}, W_{\text{proj}}\)) aggregated via weighted FedAvg, and client-local adapter + classifier heads that never leave the client.

Key Contributions

  • VQC-based multimodal fusion module using \(N=8\) qubits and \(L=3\) entangling layers, reducing fusion rotation parameters from 33K (MLP) to 72, with ~10× total reduction including \(W_{\text{pre}}/W_{\text{post}}\).
  • Personalized FL architecture that globally aggregates only encoders and VQC weights while keeping client-specific adapter and classification heads local.
  • Gradient optimization via the parameter-shift rule, enabling unbiased gradient estimates compatible with classical backpropagation.
  • Empirical demonstration that quantum entanglement achieves competitive or superior cross-modal fusion under non-IID federated training.

Results

  • QFedAgent achieves 97.7% mean test accuracy (S1: 97.8%, S2: 98.3%, S3: 97.5%, S4: 97.1%) on OPPORTUNITY dataset, outperforming all baselines.
  • Baselines on same 4-subject non-IID split: Local-only 96.3%, FedAvg 96.6%, FedProx 96.7%, MLP-FL 97.2%.
  • QFedAgent final training loss 0.186 vs. MLP-FL 0.193.
  • Fusion parameter count: 72 VQC rotation params (3,144 total including projections) vs. 33K for all classical baselines.
  • Per-round training time: 121.4 s (QFedAgent, simulated on RTX 5080) vs. 9.8–12.5 s for classical methods — a ~10× slowdown from classical simulation of the \(2^N\)-dimensional quantum state.
  • Ablation: accuracy improves with \(N\) and \(L\) up to \(N=8, L=3\); further increases yield diminishing returns with harder optimization.

Limitations

  • Classical simulation of the VQC incurs ~10× per-round training overhead (121.4 s vs. ≤12.5 s), making current deployment impractical at scale; real quantum hardware would eliminate this cost.
  • Evaluated on a single small dataset (OPPORTUNITY, 4 subjects/clients); generalization to larger or more heterogeneous federated populations is unverified.
  • Only 4 locomotion activity classes tested; performance on finer-grained or more complex activity taxonomies is unknown.
  • Subject S4's lower accuracy tied to smaller dataset size (2,780 windows), indicating sensitivity to severe data imbalance across clients.
  • No communication cost analysis is provided despite that being a stated motivation.

Relevance to Agentic AI / LLM Agents

This work is relevant to multi-agent systems research at the intersection of distributed learning and edge robotics — agents must collaboratively learn shared representations without centralizing private sensory data, a core challenge in embodied multi-agent coordination. The personalized FL design (shared global encoder, private local adapter) directly addresses the heterogeneous-agent problem where each agent has a distinct distribution over its environment. While not LLM-based, the framework illustrates a parameter-efficiency technique (quantum circuit fusion) applicable to any multi-agent architecture where communication bandwidth constrains model update size. It provides a concrete existence proof that non-classical parameterizations can match or beat classical fusion in federated agentic sensing pipelines.