Perfect Secrecy under Deep Random assumption

We present a new idea to design perfectly secure information exchange protocol, based on so called Deep Randomness, which means randomness relying on hidden probability distribution. Such idea drives us to introduce a new axiom in probability theory, thanks to which we can design a protocol, beyond Shannon limit, enabling two legitimate partners, sharing originally no common private information, to exchange secret information with accuracy as close as desired from perfection, and knowledge as close as desired from zero by any unlimitedly powered opponent.

💡 Research Summary

The paper introduces a novel concept called “Deep Randomness” and claims that it enables two legitimate parties, who share no prior secret, to exchange information with near‑perfect secrecy over a completely public channel, even against an unlimited‑power passive adversary. The authors begin by revisiting the classic information‑theoretic notion of perfect secrecy as defined by Shannon, emphasizing that its feasibility traditionally requires the secret key to have entropy at least as large as the message. They argue that this limitation stems from the implicit assumption that all parties know the prior probability distribution of the secret variables, which permits optimal Bayesian inference.

To break this assumption, the authors invoke prior‑probability theory, especially the work of Jaynes on maximum‑entropy inference. They define a “Deep Random Assumption” (DRA): each party generates its private random data according to a probability distribution that is deliberately hidden from any external observer, including the other legitimate party and the adversary. Formally, the set of admissible distributions is invariant under a finite group of transformations; the observer, lacking any reason to favor one distribution over another, must assign a uniform prior over this set. Consequently, the adversary’s Bayesian update, based on all publicly exchanged messages, cannot significantly reduce the entropy of the secret.

Based on DRA, the authors propose a two‑stage protocol. In the first stage, Alice and Bob independently generate deep‑random strings and broadcast them over a discrete memoryless main channel. In the second stage they exchange partial information (e.g., hashes or selected bits) to perform advantage distillation and error correction, ultimately extracting a shared secret key K. By tuning protocol parameters (hash length, number of rounds, etc.), the authors claim the adversary’s conditional entropy H(S|E) can be made arbitrarily close to the unconditional entropy H(S), while the legitimate receiver’s conditional entropy H(S|R) is substantially lower, satisfying H(S|E) ≥ H(S|R). Using the Csiszár‑Körner secrecy‑capacity bound, they argue that the secrecy rate C_s = H(S|E) − H(S|R) can be made positive, thereby achieving information‑theoretic security beyond Shannon’s limit.

The paper compares this approach to BB84 (quantum key distribution) and Maurer’s protocols that rely on noisy or partially independent side channels. While BB84 exploits physical quantum uncertainty and the no‑cloning theorem, and Maurer’s schemes depend on channel noise, Deep Randomness relies purely on epistemic uncertainty about the underlying distribution. The authors claim this eliminates the need for specialized hardware or physical assumptions.

A section on “generation of Deep Randomness” suggests that classical computers can produce such hidden distributions by embedding secret parameters into nonlinear functions and seeding them with physical noise (e.g., voltage fluctuations). However, the description lacks concrete algorithms, complexity analysis, or proofs that an adversary with unlimited observations cannot eventually infer the hidden distribution.

Critically, the security proof hinges entirely on the Deep Random Assumption being valid in practice. The paper does not demonstrate how to guarantee that the adversary cannot learn the hidden distribution, nor does it address side‑channel leakage, long‑term statistical learning, or composability of the protocol. Moreover, the proposed generation method appears speculative; without rigorous analysis, it is unclear whether it truly provides the required epistemic opacity.

In conclusion, the work offers an intriguing theoretical perspective: by denying the adversary knowledge of the prior distribution, one can sidestep Shannon’s classic impossibility result. Nevertheless, the practical feasibility of maintaining such hidden distributions, the robustness of the protocol against adaptive statistical attacks, and the lack of concrete implementation details leave the claim of “perfect secrecy beyond Shannon” unsubstantiated. Future research would need to provide explicit constructions of deep‑random generators, empirical validation of their secrecy properties, and a thorough analysis of how realistic adversaries might erode the assumed epistemic uncertainty.