Adversarial Deep Learning for Over-the-Air Spectrum Poisoning Attacks

An adversarial deep learning approach is presented to launch over-the-air spectrum poisoning attacks. A transmitter applies on its sensing results predict idle time slots for data transmission. In the meantime, an adversary learns transmitter's behavior (exploratory attack) by building another neural network when transmissions will succeed. The falsifies (poisons) over air transmitting during short period of transmitter. Depending whether uses as test make transmit decisions or training retrain network, either it fooled into making incorrect (evasion algorithm retrained incorrectly future (causative attack). Both attacks are energy efficient and hard detect (stealth) compared jamming long transmission period, substantially reduce throughput. dynamic defense designed that deliberately makes a small number (selected confidence score channel classification) manipulate adversary's data. This effectively fools (if any) helps sustain throughput with without present.