Adapting CoNoCo to deterministic policies

Develop adaptations of Colored Noise Coherency (CoNoCo) for deterministic robotic control policies by identifying and integrating alternative sources of variability that enable embedding a frequency-band watermark into the actions and recovering it via spectral coherency from remote observations.

Background

CoNoCo’s watermarking strategy leverages the inherent stochasticity of common continuous-control policies (Gaussian exploration noise) to embed a colored-noise signal detectable through spectral coherency. This preserves the marginal action distribution while enabling remote detection.

For deterministic policies, the stochastic channel used by CoNoCo is absent. The authors state that applying the method would require alternative sources of variability and explicitly leave this extension for future work, highlighting the need to redesign the watermarking mechanism to function without policy-driven randomness.