This paper explores image self-similarity as a means to impede forensic camera identification based on sensor noise. We follow the tradition of patch replacement attacks against robust digital watermarks, putting particular emphasis on the use of PatchMatch, an efficient algorithm for finding approximate nearest neighbor patches. Experimental results suggest that sensor noise fingerprints can be desynchronized reasonably well without the need of geometric image transformations and without explicit knowledge of the fingerprint, while maintaining a tolerable visual quality.