The hash function is an efficient source of the integrity and authentication of input text and other data messages (image & audio-video) in the cryptography field. Existing hashing algorithms are time-consuming and vulnerable to collision attacks, birthday attacks, meet-in-the-middle attacks, and key search attacks. In this context, we propose an innovative and secure hash algorithm that uses self-learning neurons during the implementation phase. Firstly, the proposed deep learning hashing algorithm accepts the variable length (ln) of the input text and image messages (im) at the input layer 0 and divides the whole input (M) into four characters, represented as message pair (MP). If the last MP contains less than four characters, have been completed by the assumption of the special cases and processed via the OR logical operation. Subsequently, the proposed pattern matching-swapping methods have resolved the conflict of the repeated characters in each MP. Secondly, the proposed novel dynamic secret keys are generated for each MP by introducing a skip and select basic conditions by omitting the consecutive 0's or 1's in key selection. Thirdly, upon successfully processing the phases above, the proposed novel hashing algorithm generates collision-free hash values using forwarding and backward propagation. Experimental results show that our proposed hashing algorithm is efficient and significantly outperforms in terms of sensitivity generated in hash output, speed, and collision resistance compared to the existing state-of-the-art hashing algorithms.