Deep Learning for Blind HF Communications Protocol Identification

Keywords

HF Signal Processing, Cybersecurity, Cognitive Radio, Machine Learning, Image processing, Acoustic analysis

Suitable Departments

Computer Science, Communications Engineering, Cybersecurity, Electronics Engineering

Summary

Cryptanalysis of blind signals is an important problem in both civilian and non-civilian domains. Resolving the protocol used in an interception determines or materially limits the set of modulation types, pilot tones frequencies, signal bandwidth, error correction approaches, compression and even encryption methodologies that were used. Once transmission parameters are fully identified, data packets can be correctly identified in the transmission. This is a necessary but insufficient condition for cryptanalysis.

State-of-the-art systems jointly determine transmission parameters and the message content. The process is computationally demanding except in the case of cooperative communications in which the parameter sets are exchanged or communicated a priori using a secure, pre-agreed protocol. We develop an efficient approach for classifying blind signals using a hierarchical machine learning approach wherein computational complexity of classifiers is progressively increased depending on residual entropy. At one end we have regression or tree-based approaches. At the computationally demanding end is deep learning (DL) that works with spectral features, including Haar-like features.

The novelty of this work is an algorithm for automatically pruning and rationalizing bit-quantization to arrive at a much smaller network for classifying blind signals without material compromises on accuracy. The size of the maximal DL network depends on the computational resource limitations of the particular hardware. We will address the challenge of accommodating hardware with different specifications without the need to retrain.

Programming Skills Needed

Matlab/Python, C++

Hardware Skills Needed

An understanding of computer architecture and hands-on experience with Arduino or Raspberry Pi is helpful though not necessary

• We will not intercept the transmissions. Datasets from online repositories will be used.
• We will not implement on a chip. We will instead emulate their limitations on a PC.

Other Skills

Parallel Computing, Algorithm Design, Machine Learning, Communications Theory, Image Processing, Acoustic Signal Processing

Partners

India : industry
Outside : Dr. Vasily Sachnev

Funding

Amrita Vishwa Vidyapeetham provides stipends and teaching assistantships to selected candidates. Once you join, you could leverage existing proposals within the team to apply for additional corporate or government funding.

Time Period

3 ½ to 4 ½ years based on full-time committed work.