To make their analytical service reliable, the distributed machine learning solutions must secure inter-device data transmission from several cyber threats such as jamming, eavesdropping, denial of service, and so on. However, data transmission among the distributed, densely deployed machine learning agents are more vulnerable, compared to the traditional communication equipment, to cyber-attacks for the following reasons:
- With their pervasive existence, information transmitted among the distributed agents are more likely to be intercepted by undesired receivers (i.e., eavesdroppers) and, thus, cause information leakage.
- It is rather challenging to design security schemes for satellites and UAVs with their limited computation ability and energy resource. Yet the exposed inter-device communication can cause major security issues for distributed machine learning solutions, as the the leakage of these information can help an intelligent attacker resume the original data in the system.
Thus, a key goal here is to develop novel security solutions considering the limited resource capacity of the distributed machine learning systems, such solutions include:
- Privacy-reserving meta federated learning-based spectrum management, beam training, and positioning solutions at the integrated system to effectively avoid jamming attacks with minimal training cost. It is of interest to develop algorithms that can train wireless strategies distributed at each interdependent device in the system, to reduce overhead, and complexity of the considered connectivity optimization problem.
- Artificial noise intelligently carried out by UAVs to reduce the capacity of eavesdropping channels. We will propose task-robust meta learning solutions to quickly adapt UAVs in various unknown environments, which includes expected environments from certain distribution and the ones out of the expected distribution. The proposed solution will adjust UAV learning process within the task of reaching a noise launching scheme effectively interfering every possible eavesdropper from unknown locations.
- 3D beamforming to reduce the possibilities of information leakage. Specifically, we will use a meta learning-based solution to predict the mobility patterns of ground users for effective beamforming design.