RL-PDNN: Reinforcement Learning for Privacy-Aware Distributed Neural Networks in IoT Systems

Abstract

Due to their high computational and memory demand, deep learning applications are mainly restricted to high-performance units, e.g., cloud and edge servers. Particularly, in Internet of Things (IoT) systems, the data acquired by pervasive devices is sent to the computing servers for classification. However, this approach might not be always possible because of the limited bandwidth and the privacy issues. Furthermore, it presents uncertainty in terms of latency because of the unstable remote connectivity. To support resource and delay requirements of such paradigm, joint and real-time deep co-inference framework with IoT synergy was introduced. However, scheduling the distributed, dynamic and real-time Deep Neural Network (DNN) inference requests among resource-constrained devices has not been well explored in the literature. Additionally, the distribution of DNN has drawn the attention to the privacy protection of sensitive data. In this context, various threats have been presented, including white-box attacks, where malicious devices can accurately recover received inputs if the DNN model is fully exposed to participants. In this paper, we introduce a methodology aiming at distributing the DNN tasks onto the resource-constrained devices of the IoT system, while avoiding to reveal the model to participants. We formulate such an approach as an optimization problem, where we establish a trade-off between the latency of co-inference, the privacy of the data, and the limited resources of devices. Next, due to the NP-hardness of the problem, we shape our approach as a reinforcement learning design adequate for real-time applications and highly dynamic systems, namely RL-PDNN. Our system proved its ability to outperform existing static approaches and achieve close results compared to the optimal solution. INDEX TERMS IoT devices, distributed DNN, privacy, white-box, resource constraints, DQN.

Optimization Framework for Splitting DNN Inference Jobs over Computing Networks

“…Proof of Theorem 2: By Lemmas 6 and 5, A is TU. By Lemmas 2 and 1, LP relaxation (7) has an integral optimal solution. It is clear that LP relaxation ( 7) is equivalent to (13).…”

“…The slack variable s 2 is introduced to replace the inequality constraint with equality constraint again, i.e., replace y ≤ 1 with y +s 2 = 1 and s 2 ≥ 0. Putting all the equality constraints together obtains the desired result (7). Note that the polyhedron in LP relaxation (7) is the same form as in Lemma 1 with equality Ax = b, where…”

“…Putting all the equality constraints together obtains the desired result (7). Note that the polyhedron in LP relaxation (7) is the same form as in Lemma 1 with equality Ax = b, where…”

“…The goal is to find a mapping that minimizes the network-wide total transmit power or time while the computing resource needed for assigned neurons and daily energy consumption satisfy the specified limit. The work of [7] assumes layer-wise partition, and seeks to find a mapping that minimizes (data) transmission time plus computation time, subject to privacy constraints as well as various resources constraints. This problem is formulated as an integer quadratic program, and a reinforcement learning based mapping algorithm is proposed.…”

Pervasive AI for IoT Applications: Resource-efficient Distributed Artificial Intelligence

“…More specifically, the distribution system has to account for the past tasks assigned to each participant and try to maximize the re-usability of previously-stored weights, with consideration to the capacity of the devices. Minimizing the number of weights assigned to each participant, not only contributes to reduce the memory usage, but also guarantees the privacy of the structure against white-box attacks [307].…”