This paper introduces a concept of Active Content Fingerprinting based on a Latent data Representation (aCFP-LR). The idea is to represent the data content by a constrained redundant description. The target is to estimate latent representation such that: (i) after applying a reconstructor function the result is close to the original data and (ii) after using an extraction function the resulting features are robust. A general problem formulation is proposed for aCFP-LR with an extractor-reconstructor pair of constraints. One particular case is considered under linear extractor (generator) and linear reconstructor (modulator) where a reduction is shown to a constrained projection problem. Evaluation by numerical experiments is given using local image patches, extracted from publicly available data sets. Advantages and state-of-the-art performance is demonstrated under additive white Gaussian noise (AWGN), lossy JPEG compression and projective geometrical transform distortions.