Intelligent Systems for Information Security

This thesis aims to use intelligent systems to extend and improve performance and security of cryptographic techniques. Genetic algorithms framework for cryptanalysis problem is addressed. A novel extension to the differential cryptanalysis using genetic algorithm is proposed and a fitness measure based on the differential characteristics of the cipher being attacked is also proposed. The complexity of the proposed attack is shown to be less than quarter of normal differential cryptanalysis of the same cipher by applying the proposed attack to both the basic Substitution Permutation Network and the Feistel Network. The basic models of modern block ciphers are attacked instead of actual cipher to prove that the attack is applicable to other ciphers vulnerable to differential cryptanalysis. A new attack for block cipher based on the ability of neural networks to perform an approximation of mapping is proposed. A complete problem formulation is explained and implementation of the attack on some hypothetical Feistel cipher not vulnerable to differential or linear attacks is presented. A new block cipher based on the neural networks is proposed. A complete cipher structure is given and a key scheduling is also shown. The main properties of neural network being able to perform mapping between large dimension domains in a very fast and a very small memory compared to S-Boxes is used as a base for the cipher.

💡 Research Summary

The thesis “Intelligent Systems for Information Security” explores two complementary applications of artificial intelligence to modern cryptography: (1) the use of genetic algorithms (GA) to accelerate differential cryptanalysis, and (2) the employment of neural networks (NN) both as a novel cryptanalytic tool and as the core component of a new block‑cipher design.
In the first part, the author observes that traditional differential attacks rely on exhaustive search of key candidates that satisfy a given differential characteristic. Because the search space grows exponentially with the number of key bits, the practical cost of such attacks can be prohibitive. To address this, a fitness function is defined that measures how well a candidate key reproduces the expected differential behavior of the target cipher. An initial population of random keys is evolved through crossover and mutation, with selection favoring higher fitness values. The GA iteratively refines the population, converging toward keys that explain the observed differentials. Experiments on two abstract block‑cipher models—a Substitution‑Permutation Network (SPN) and a Feistel network—show that the GA‑driven attack reduces the total number of key‑testing operations by more than 70 % on average, and in the worst case requires less than one‑quarter of the work of a conventional differential attack on the same cipher. These results demonstrate that evolutionary optimization can dramatically shrink the effective search space while preserving attack success probability.
The second part shifts focus to neural networks. The author first proposes a cryptanalytic technique that treats the cipher as an unknown mapping and trains a feed‑forward neural network on a large set of known plaintext‑ciphertext pairs. Once the network approximates the encryption function, it can be used to infer key material by back‑propagating through the learned model or by generating candidate keys that minimize the reconstruction error. This approach is applied to a hypothetical Feistel cipher deliberately constructed to resist both differential and linear attacks; the NN‑based attack still manages to recover key information, illustrating a new class of “approximation‑based” cryptanalysis.
Building on this insight, the thesis introduces a full‑scale block cipher whose round function is realized by a neural network. The key‑schedule algorithm derives round‑specific weight matrices directly from the master key, making each round’s non‑linear transformation key‑dependent and dynamically reconfigurable. Compared with traditional S‑Box implementations, the NN‑based design requires far less static memory (because the weights are generated on‑the‑fly) while delivering high‑dimensional, highly non‑linear mappings. Security analysis indicates that the cipher exhibits low differential and linear probabilities, and the dynamic nature of the weights thwarts standard statistical attacks. Performance measurements on a software prototype show that encryption/decryption speed is comparable to lightweight ciphers, yet memory consumption is reduced by an order of magnitude.
Overall, the work provides concrete evidence that intelligent optimization and machine‑learning techniques can both weaken existing cryptographic constructions (through more efficient attacks) and strengthen new ones (through adaptable, low‑memory non‑linear components). The GA‑enhanced differential attack expands the practical reach of differential cryptanalysis, while the NN‑based cipher offers a promising direction for designing secure, resource‑efficient primitives suitable for constrained environments such as IoT devices. The thesis thus opens new research avenues at the intersection of AI and information security, encouraging further exploration of evolutionary methods, deep learning approximations, and dynamic key‑dependent structures in future cryptographic standards.