Introduction
Modern autonomous and connected vehicles generate vast amounts of sensor data, often exceeding the local processing capabilities of onboard units. While offloading to the cloud offers a solution, it introduces significant privacy risks and network congestion. Traditional Information Bottleneck (IB) methods provide efficient data compression but are computationally heavy for single vehicles. This paper introduces BVIB (Blockchain-Enabled Variational Information Bottleneck), a framework that balances computational efficiency with high security for the Internet of Vehicles (IoV).
Methodology
The BVIB framework utilizes a synergistic approach combining deep learning model splitting with distributed ledger technology.
Model Splitting and Blockchain Integration
Instead of executing a full Variational Information Bottleneck (VIB) model locally, the architecture is segmented. The Encoder, responsible for compressing raw data into a compact latent representation, remains on the vehicle. The Decoder, which reconstructs or interprets the data, is offloaded to a blockchain-secured server.
The servers operate as nodes in a blockchain network utilizing the Raft consensus algorithm. This ensures data immutability and rapid recovery through leader election if a node is compromised.
Latency and Rate Optimization
The system accounts for various delays, including encoding, wireless transmission failures (modeled via a Markov process), and blockchain overhead. By applying convex optimization, the authors identify an optimal data extraction rate ($\lambda^*$) that minimizes total network latency, effectively managing the trade-off between data throughput and channel collisions.
Experimental Evaluation
The framework was evaluated using the MNIST dataset. While a standard image benchmark, it serves as a proof-of-concept for feature compression and transmission under varying network conditions and adversarial attacks.
Key Insights
- Efficiency Gains: Offloading the decoder reduced the vehicle's neuron count by approximately 36%, leading to a corresponding 36% reduction in time cost compared to standard VIB or VAE approaches.
- Mutual Information Filtering: By optimizing for Mutual Information, BVIB discards non-essential "noise," making transmitted data more robust against interference.
- The Latency "Sweet Spot": The research identifies a mathematical "U-curve" for latency relative to the extraction rate. Finding the bottom of this curve in real-time prevents both idle waiting and excessive network collisions.
Results and Security Performance
BVIB demonstrates superior resilience against malicious actors. In simulations involving multiple malicious nodes, BVIB and Blockchain Federated Learning (BFL) maintained significantly higher accuracy than non-blockchain methods like VAE and VIB, which suffered sharp performance declines.
Furthermore, BVIB achieved lower latency than BFL because it transmits compressed latent vectors rather than entire model parameters. Crucially, the "split" architecture does not degrade core performance, as BVIB achieved nearly identical compression accuracy to the original centralized VIB.
Conclusion and Future Directions
BVIB successfully addresses the "heavy compute vs. high security" dilemma in IoV. Future work will focus on scaling the framework for high-definition video and LiDAR data, managing high-speed mobility handovers, and optimizing transmission spacing in hyper-dense urban environments.
Summary updated