The looming threat of quantum computing, particularly Shor's algorithm, to widely-used encryption algorithms such as RSA is now a sword of Damocles hanging over the security of modern information systems—a longstanding danger to the RSA algorithm and the entire security system. Against this backdrop, and motivated by advances in quantum cryptography and large AI models,this paper proposes a new encryption paradigm based on large models. In this framework,training data and information are defined as the plaintext, the training process is the encryption, forward reasoning is decryption, and the training parameters —including the training data format, process parameters, and initial parameters—are all considered encryption keys, the weights of the model can be mapped to ciphertext's probability density distribution, and the prompt serves as decryption key. The security of this encryption system primarily stems from four aspects: (i) the absence of first principles underlying the current large model's data compression activation path distribution (e.g., in transformers), (ii)the structural security of neural network's high-dimensional space distributed storage (hiding a leaf in the forest), (iii) the resistance to adversarial attacks brought by probabilistic decryption, and (iv) the existence of high-dimensional crevices. Analysis shows that this encryption method does not rely on traditional mathematical puzzles, can withstand quantum computing attacks, and provides a new research direction and theoretical basis for post-quantum cryptography.
Lihan Xia (Wed,) studied this question.