Spike Neural Network Study: Covert Signals Escape by Moving, Not Hiding

IgorPereverzevDev, an AI safety researcher transitioning from software engineering, conducted experiments on adversarial steganography using SpikeGPT, a 216M-parameter RWKV-based spiking neural network that processes information via discrete spikes. The setup had Alice attempting to transmit a secret signal to Bob while Mallory (a monitor) observed the channel. Over fourteen rounds of joint training, the signal was encoded and the monitor retrained to catch up. Surprisingly, the hidden signal remained trivially readable by a 45-parameter linear probe with 100% accuracy at every checkpoint—yet Mallory's detection accuracy plummeted from 100% to around 30%, and in some rounds below 21%.

The explanation lies in geometric translation: the signal lived as clusters of points in latent space. Alice did not obfuscate the clusters or mix their contents; she moved them as whole entities. The cluster centroids shifted by an average of 2.5 units, while the internal spread within each cluster changed by only 0.26 units—a ratio of 9.7:1. The linear probe, retrained each round, followed the clusters to their new locations, while the monitor, trained on previous locations, faced geometric shift. This implies that covert channels can evade detection through movement, not complexity. For AI safety monitoring, robustness to distribution shift may matter more than pure discriminatory ability.