Encrypted data is often identifiable (e.g., k*m2), as its features differ from those of unencrypted data (e.g., home). This paper proposes a novel solution for this problem: Context Preserving Encryption (CPE). By extracting features from data and encrypting those features, then generating new data from the encrypted features, CPE can create ciphertext (e.g., the word “bicycle” or realistic face images) that is indistinguishable from plaintext. To generalize CPE to continuous real-world applications, I introduce a novel neural network-based encryption regime that inverts the autoencoder architecture to reconstruct latent vectors while the middle layer representations are enforced to resemble real data. Simulation with handwritten digits and facial data demonstrated the effectiveness of neural CPE. Applying CPE can help reduce the risk of becoming a target of potential attackers by maintaining the secrecy of the existence of a secret, and give plausible deniability if detected.