AI-Powered Cryptanalysis: Identifying Encryption Algorithms and Recovering Plaintext

With encryption becoming more prevalent for the security of digital correspondence, the actual process of analyzing the ciphertext without the decryption key becomes one of the single biggest problems in cybersecurity and cryptanalysis. This represents two fundamental problems: classifying ciphertext based on the encryption scheme used, and reconstructing plaintext from encrypted sequences leveraging deep learning. The more classic style approaches to cryptanalysis often rely on brute force or some mathematical “weakness” in the algorithm itself, but with the advent of neural networks, the cryptanalysts are able to discover patterns to the structural data represented in the encrypted data. This paper deploys the bidirectional long short-term memory (BiLSTM) and bidirectional gated recurrent unit (BiGRU) neural networks to classify ciphertext produced by the Advanced Encryption Standard (AES), Triple Data Encryption Standard (3DES), Blowfish, and Twofish encryption schemes into the respective categories. The BiLSTM model was able to classify the ciphertext with a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 7. 9 1 \%}$</tex> classification accuracy for the dataset, with the 1.07 % better performance over the BiGRU model, which successfully classified the dataset with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 6. 9 8 \%}$</tex> accuracy. The second part of the research involved the use of a sequence-to-sequence long short-term model to reconstruct original text from ciphertexts encrypted under the Data Encryption Standard (DES) and Twofish - plaintext was provided from the Internet Movie Database (IMDB) dataset. The reconstruction accuracy of DES-encrypted ciphertext was high, achieving an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex>-score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 8 6 8}$</tex>, which supports that certain encryption schemes may retain exploitable patterns on which deep learning models can be trained. In contrasting examples, the Twofish-decrypted ciphertext was lowered to an F1-score of 0.750 resulting in a lower <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex> by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 3. 6 \%}$</tex> due to heavier diffusion which produced additional resistance. The above findings demonstrate the efficacy of neural models to detect and exploit structural weaknesses in legacy encryption systems and call for encryption algorithms to reduce recoverable features against deep learning attacks. The study provides the first step for future studies involving artificial intelligence driven tools assisting in forensic cryptography, automated vulnerability assessment, and secure system design.