On Learning-Assisted Content-Based Secure Image Transmission for Delay-Aware Systems With Randomly-Distributed Eavesdroppers

Mehdi Letafati, Hamid Behroozi, Babak Hossein Khalaj, Eduard A. Jorswieck

TL;DR

This paper proposes a novel learning-assisted wireless image transmission scheme that prioritizes image content for security and adapts transmission parameters to minimize Quality-of-Service (QoS) violations, specifically the probability of eavesdroppers obtaining sensitive image regions before the legitimate receiver. By leveraging deep neural networks (DNNs), the system achieves a balance between security, delay, and reliability, outperforming existing methods.

What Problem Does This Paper Address?

The increasing demand for high-quality wireless services like AR/VR and surveillance systems necessitates secure and reliable image transmission. However, existing wireless systems face significant security challenges, especially with the proliferation of low-cost eavesdropping devices. Furthermore, not all parts of an image are equally important from a security perspective. Transmitting entire images with uniform security can be inefficient and lead to delays. This paper addresses the challenge of providing content-aware, delay-aware, and secure wireless image transmission in the presence of multiple, randomly distributed passive eavesdroppers. It aims to ensure that sensitive image regions are protected while meeting strict delay requirements, a problem not fully addressed by prior work that often treats all data equally or focuses solely on perfect secrecy without considering delay constraints.

Methods

The proposed scheme employs a hybrid approach:

  • Content-Aware Segmentation: Images are divided into "regions of interest" (RoI) containing important information and "background" (BG) regions with less critical data.
  • Hybrid Transmission Policy:
    • RoI Transmission: Utilizes Maximum Ratio Transmission (MRT) beamforming combined with Artificial Noise (AN) injection to enhance the legitimate link's signal and degrade eavesdroppers' capabilities.
  • BG Transmission: Employs MRT beamforming alone to conserve power and improve packet accumulation at the receiver, as BG regions require less stringent security.
  • Fountain Coding (FC): Source packets are encoded into fountain-like packets to enable efficient packet reconstruction at the receiver, even with packet loss.
  • Delay-Awareness: The system considers delay limits (Dlim) and adapts the number of packets transmitted per time slot (TS) to meet these constraints.
  • Learning-Assisted Optimization: A Deep Neural Network (DNN) is trained offline to learn the optimal transmission parameters (e.g., power allocation, AN ratio, decision threshold) that minimize the Quality-of-Service Violation Probability (QVP). This DNN maps system configurations (image size, channel conditions, receiver location) to optimal parameters, enabling efficient real-time adaptation.
  • Random Eavesdropper Model: Eavesdroppers are modeled as a homogeneous Poisson Point Process (PPP), considering both non-colluding (NCE) and colluding (CE) scenarios.

Datasets

Not discussed in the paper. The paper describes generating a training dataset using numerical methods (genetic algorithm) to solve the optimization problem offline.

Key Insights and Ideas

  • Content is King for Security: Prioritizing security for RoI packets over BG packets is crucial for efficient and effective secure image transmission, especially in delay-constrained scenarios.
  • AN is Essential for Multiple Eavesdroppers: While adaptive multi-packet transmission helps, Artificial Noise injection is vital for effectively combating multiple passive eavesdroppers, particularly in the colluding (CE) scenario.
  • DNNs as Efficient Optimizers: Deep Neural Networks provide an efficient and flexible way to find optimal transmission parameters in complex, non-linear wireless environments, significantly reducing computational complexity compared to traditional iterative methods, especially for real-time adaptation.
  • Delay vs. Security Trade-off: The QVP metric effectively captures the trade-off between timely delivery and security, showing how increasing delay tolerance can improve security by allowing more packets to be accumulated.
  • Channel Estimation Quality Matters: Accurate estimation of the legitimate channel significantly improves system performance (lower QVP) by enabling better beamforming and reducing the impact of AN on the legitimate link.

Key Results

  • The proposed learning-assisted scheme achieves significantly lower QVP compared to benchmark schemes (AN-aided Fixed Rate and AN-free Adaptive Rate) across various delay limits.
  • The system demonstrates the ability to achieve arbitrarily small File Intercept Probabilities (FIP) by adjusting transmission parameters, particularly by increasing the number of transmit antennas.
  • The DNN-based optimization effectively tunes parameters to minimize QVP, outperforming fixed or equal power allocation strategies.
  • Numerical results show good agreement between analytical derivations and simulation results, validating the proposed models and performance metrics.
  • The paper demonstrates that increasing the number of transmit antennas leads to lower FIP due to improved beamforming.

Limitations and Future Work

  • Image Segmentation: The paper assumes image segmentation is performed but does not delve into the methods for achieving this, stating it's out of scope.
  • Computational Complexity of Training: While the online phase is computationally efficient, the offline training phase, especially data set generation, can be computationally intensive.
  • Generalization of DNN: The paper focuses on a specific system configuration for training. Future work could explore more robust DNN architectures that generalize better across a wider range of system parameters and environments.
  • Eavesdropper Sophistication: The model assumes passive eavesdroppers. Future work could consider more active eavesdropping strategies.
  • Real-world Deployment: While the paper presents a theoretical framework and simulation results, real-world implementation challenges and performance would be an interesting area for future research.

Summary updated

Abstract

In this paper, a learning-aided content-based image transmission scheme is proposed, where a multi-antenna source wishes to securely deliver an image to a legitimate destination in the presence of randomly-distributed passive eavesdroppers (Eves). We take into account the fact that not all regions of an image have the same importance from the security perspective. Hence, we employ a hybrid method to realize both the error-free data delivery of public regions—containing less-important pixels; and an artificial noise (AN)-aided transmission scheme for securing the confidential packets. To reinforce system’s security, fountain-based packet delivery is also adopted, where the source node encodes images into fountain-like packets prior to sending them over the air. The secrecy is achieved when the legitimate destination correctly receives the entire source packets before Eves obtain the important regions, while conforming to the latency limits of the system. Accordingly, the secrecy performance of our scheme is characterized by deriving a closed-form expression for the quality-of-security (QoSec) violation probability. Moreover, our proposed image delivery scheme leverages a deep neural network (DNN) and learns to maintain optimized transmission parameters, while achieving a low QoSec violation probability. Simulation results are provided to illustrate that our proposed learning-assisted scheme outperforms the state-of-the-arts by achieving considerable gains in terms of security and delay requirement.
BibTeX
@article{letafati2022learning,
  author = {Mehdi Letafati and Hamid Behroozi and Babak Hossein Khalaj and Eduard A. Jorswieck},
  title = {{On Learning-Assisted Content-Based Secure Image Transmission for Delay-Aware Systems With Randomly-Distributed Eavesdroppers}},
  journal = {IEEE Trans. Communications},
  year = {2022},
  doi = {10.1109/tcomm.2021.3128423},
}
Figure