Federated Distillation Assisted Vehicle Edge Caching Scheme Based on Lightweight DDPM

Vehicle edge caching is a promising technology that can significantly reduce the latency for vehicle users (VUs) to access content by pre-caching user-interested content at edge nodes. It is crucial to accurately predict the content that VUs are interested in without exposing their privacy. Traditional federated learning (FL) can protect user privacy by sharing models rather than raw data. However, the training of FL requires frequent model transmission, which can result in significant communication overhead. Additionally, vehicles may leave the road side unit (RSU) coverage area before training is completed, leading to training failures. To address these issues, in this letter, we propose a federated distillation-assisted vehicle edge caching scheme based on lightweight denoising diffusion probabilistic model (LDPM). The simulation results demonstrate that the proposed vehicle edge caching scheme has good robustness to variations in vehicle speed, significantly reducing communication overhead and improving cache hit percentage.

💡 Research Summary

The paper proposes a novel vehicle edge caching framework that simultaneously addresses the challenges of high vehicle mobility, privacy preservation, and communication overhead in vehicular networks. The authors combine a lightweight Denoising Diffusion Probabilistic Model (LDPM) with a federated distillation (FD) mechanism to predict user‑interested content and pre‑cache it at roadside units (RSUs) without exposing raw user data.

System Architecture
The network consists of three layers: a macro base station (MBS), multiple RSUs, and vehicles. Each vehicle holds local interaction data and is equipped with a pre‑trained encoder‑decoder pair. When a vehicle enters an RSU’s coverage, it encodes its data into a compact hash code using the encoder and uploads this hash together with its identifier (HI pair). The RSU stores the hash in a hash‑to‑index (HI) module and uses cosine similarity to find the C most similar neighboring vehicles. Knowledge vectors (outputs of the local LDPM) from these neighbors are retrieved from a knowledge‑cache (KC) module, averaged to form an integrated knowledge vector ˜mᵣᵢ, and sent back to the vehicle.

Lightweight DDPM (LDPM)
Traditional DDPMs are computationally intensive; the authors adopt a lightweight U‑Net architecture that replaces 2‑D convolutions with 1‑D convolutions, reduces channel depth to one‑quarter, and thus contains only ~770 K parameters. This makes it feasible for on‑board execution. The LDPM operates in a latent space: the vehicle first maps its raw data dᵢ to a low‑dimensional representation ˆdᵢ = Eᵢ(dᵢ). The forward diffusion adds Gaussian noise, while the reverse diffusion (trained by the LDPM) denoises to generate synthetic user‑item interaction vectors.

Federated Distillation Training
Instead of transmitting the full LDPM parameters as in conventional federated learning (FL), each vehicle performs local distillation training using two loss components: (1) a standard reconstruction loss on its own latent samples, and (2) a KL‑divergence term that aligns the vehicle’s output distribution with the integrated knowledge ˜mᵣᵢ received from the RSU. The temperature‑scaled softmax (δ) and weighting factor (λ) control the influence of the knowledge term. After training, the vehicle runs the reverse diffusion to produce F synthetic interaction vectors, decodes them back to the original dimension, aggregates them to obtain a content‑preference score vector ˜gᵣᵢ, and selects the top‑M items as its recommendation list Lᵢ,ᵣ. Only the recommendation list and the knowledge‑index (KI) pair (knowledge vector + vehicle ID) are uploaded to the RSU, keeping communication payload low and preserving privacy.

Cache Management at RSU
The RSU updates its cache using a mobility‑aware scoring function: for each content k, the score uᵣ(k) = Σᵢ∈Uᵣ η·(B−Pᵢ)/Vᵣᵢ·𝟙{k∈Lᵢ,ᵣ}, where η is a location‑weight factor, B the RSU coverage length, Pᵢ the distance of vehicle i from the RSU entrance, and Vᵣᵢ the vehicle speed. The top‑N contents with highest scores are stored, reflecting both user preference and the expected dwell time of vehicles. The RSU periodically uploads its KC to the MBS; the MBS aggregates KC from all RSUs, retains the most recent KI/HI for each vehicle, and disseminates the aggregated KC back to RSUs, ensuring that knowledge remains fresh even after vehicles leave the coverage area.

Experimental Evaluation
Simulations use the MovieLens‑1M dataset (≈6 k users, ≈4 k movies) to emulate vehicle requests. Key parameters include cache capacity N=500, average vehicle speed 25 m/s, diffusion steps T=50, temperature δ=2, KL weight λ=1, and number of similar neighbors C. Baselines comprise: (i) Oracle (full prior knowledge), (ii) conventional FL‑based edge caching, and (iii) simple popularity‑based caching. Results show:

• Communication Overhead – FD reduces transmitted data by >70 % compared with FL because only hash codes (≈10 KB) and knowledge vectors are exchanged, not the full 770 K‑parameter model.
• Cache Hit Ratio – Across vehicle speeds from 10 m/s to 40 m/s, the proposed scheme improves hit ratio by 8–12 percentage points over FL and up to 15 pp over popularity‑based methods.
• Content Request Latency – Average latency drops by 15–20 % due to higher hit rates and mobility‑aware weighting.
• Training Interruption – Because FD does not require synchronized global rounds, vehicles that exit RSU coverage do not cause training failures; the interruption rate is effectively zero, demonstrating robustness to high mobility.

Contributions

1. Introduces a lightweight DDPM suitable for on‑vehicle inference and generation of high‑quality preference samples.
2. Proposes a federated distillation protocol that replaces full‑model exchanges, dramatically cutting communication cost while preserving model performance.
3. Designs a hash‑based similarity search and knowledge integration pipeline that safeguards user privacy and enhances prediction accuracy.
4. Develops a mobility‑aware cache replacement strategy and a KC synchronization mechanism to maintain up‑to‑date knowledge in dynamic vehicular environments.

Future Directions
The authors suggest extending the framework to multi‑RSU collaborative distillation, asynchronous model updates, and real‑world vehicular testbeds. Moreover, applying the approach to other data modalities (e.g., map updates, sensor streams) and exploring adaptive selection of C and diffusion steps based on network conditions are identified as promising research avenues.