Deep Learning Based Over-the-Air Training of Wireless Communication Systems without Feedback.

Area of Science:

• Wireless Communications
• Deep Learning
• Signal Processing

Background:

• Deep learning for over-the-air training in wireless systems faces challenges due to channel discontinuities.
• Existing methods rely on feedback channels, increasing resource demands and vulnerability to attacks.
• Feedback-free training on the forward link alone presents difficulties in reliably ending the training process.

Purpose of the Study:

• To propose a novel method for over-the-air training of wireless communication systems that eliminates the need for a feedback channel.
• To enable simultaneous training of transmitter and receiver models without continuous channel sounding.
• To offer a more resource-efficient and secure training paradigm.

Main Methods:

• Transmitting random samples through the channel to train a mixture density network (MDN) for channel distribution approximation.
• Utilizing the trained MDN to train both transmitter and receiver models.
• Employing block error rate (BLER) measurements as a stopping criterion during training.

Main Results:

• The proposed method successfully trains transmitter and receiver models without a feedback channel.
• Block error rate measurements proved effective for monitoring training completion.
• Achieved performance equivalent to end-to-end autoencoder training for small message sequences.

Conclusions:

• The novel feedback-free approach enables efficient and secure over-the-air training of wireless communication systems.
• This method simplifies the training process by removing the need for a feedback channel.
• The technique demonstrates comparable performance to existing methods, offering a viable alternative.