In this paper, we advocate an alternative signal processing based approach to searchable encryption architectures allowing to find non-exact or similar matches in the encrypted domain. The proposed approach is based on a modified architecture, where the main computational load is reallocated to a data user, who challenges an unsecure server by multiple requests, while the role of the server is reduced to appropriately replying to these challenges. To minimize the number of challenges per query, we propose a concept of bit reliability allowing to filter out the most reliable bits to formulate the most precise query in the shortest number of steps that can match the encrypted counterpart stored in the server database. Several practical implementations are discussed and empirical upper bounds on the search accuracy in terms of average probability of error are obtained for real image search under various distortions including additive Gaussian noise, uniform noise and lossy JPEG compression.