Core Concepts
A deep reinforcement learning-based vehicle selection scheme is proposed to enhance the safety and accuracy of asynchronous federated learning in vehicular edge computing by considering vehicle mobility, computational resources, data size, time-varying channel conditions, and Byzantine attacks.
Abstract
The paper presents a method for vehicle identification and selection for asynchronous federated learning (AFL) in vehicular edge computing (VEC) to improve the accuracy and safety of the global model.
Key highlights:
- A deep reinforcement learning (DRL) based vehicle selection scheme is proposed to mitigate the influence of vehicles with limited local data, poor computational capabilities, poor communication channel status, and those affected by Byzantine attacks on the global model aggregation.
- The vehicle's mobility, time-varying channel status, and time-varying computational capabilities are considered in the AFL aggregation process to give more weight to vehicles with better performance, enhancing the global model's precision.
- A threshold-based mechanism is introduced to identify and exclude vehicles affected by Byzantine attacks before the AFL training, preventing the attacks from degrading the global model's accuracy.
- Simulation results demonstrate that the proposed scheme effectively improves the safety and accuracy of the global model compared to traditional AFL and federated learning approaches.
Stats
The local training delay of vehicle n is calculated as:
T^n_l = D_n * C_0 / μ_n
The transmission delay of uploading the local model of vehicle n at time t is calculated as:
T^n_u(t) = |w| / R_n(t)
The transmission rate of vehicle n at time t is calculated as:
R_n(t) = B * log2(1 + p_0 * h_n(t) * (d_n(t))^(-α) / σ^2)
The channel gain h_n(t) is modeled using an auto-regressive model:
h_n(t) = ρ_n * h_n(t-1) + √(1-ρ_n^2) * e(t)
Quotes
"In vehicular networks with edge assistance, vehicle mobility has to be taken into account."
"The variation in data quantity and computation capabilities among vehicles will impact the local training time in AFL and time-varying features of channels will also affect the model upload rate."
"In AFL, vehicles' data may be attacked by Byzantine attacks. The prevention of Byzantine attacks is significantly important for AFL in VEC."