The short version: a pending patent application from Verizon Patent and Licensing Inc., published June 25, 2026 as US20260180793A1, describes generating encryption keys with a generative adversarial network. A key generator network produces candidate keys; a key discriminator network, trained against synthetic data derived from multiple key derivation functions, scores their authenticity; and the adversarial training objective is to push the generated keys toward maximum entropy. It is a published application — pending, not granted — and is classified under CPC H04L 9/0861, the subclass covering key generation.

The strength of any symmetric encryption key rests on a single property the rest of the cryptosystem cannot manufacture for it: entropy. A key that an attacker can guess, narrow, or enumerate is a weak key no matter how strong the cipher wrapped around it. Conventional systems lean on key derivation functions (KDFs) such as PBKDF2, scrypt, or HKDF to stretch and condition input material — a password, a shared secret, some sampled randomness — into a fixed-length key. KDFs are deterministic, well-studied transforms, but their output entropy is ultimately bounded by the entropy of what you feed them. The Verizon application, titled "Systems And Methods For Utilizing Machine Learning Models To Generate Encryption Keys," is directed at that ceiling: it frames key generation not as a fixed transform but as a learned, adversarial process whose stated goal is keys with higher entropy than the KDFs it generalizes.

How the disclosed approach works

The architecture is a generative adversarial network, the same two-network setup that underlies image and audio synthesis, applied here to bit strings that will be used as keys. The application describes the high-level behavior in its abstract:

A device may receive input key material, and may process the input key material, with a trained generative adversarial network (GAN) model, to generate an encryption key with a maximized entropy. The trained GAN model may include a key generator network model trained to generate encryption keys that generalize key derivation functions with higher entropy to enhance cryptographic security, and a key discriminator network model trained to predict authenticities of the encryption keys generated by the key generator network model.— Systems And Methods For Utilizing Machine Learning Models To Generate Encryption Keys, US20260180793A1

Read against independent claim 1, the mechanism becomes more concrete. A device receives input key material and multiple key derivation functions, then uses those KDFs to generate what the claim calls "synthetic key derivation function data." That synthetic data is the discriminator's training target: the key discriminator network is trained to distinguish the synthetic KDF data from the keys the generator produces. In a standard GAN the discriminator learns to tell "real" samples from "fake" ones, and the generator improves by learning to fool it. Here the adversarial pressure is bent toward a cryptographic objective — the claim states the discriminator is trained "to maximize an entropy of the encryption key." In other words, the generator is rewarded not just for producing outputs that resemble KDF output, but for producing outputs the discriminator cannot cleanly separate from a maximally unpredictable distribution.

Several dependent claims sketch how the system is tuned and measured. The application describes incorporating noise into the input key material before processing, and specifies in one dependent claim that the noise may include thermal noise, shot noise, or impulse noise — physical entropy sources folded into the generator's input. Other dependent claims describe calculating a loss function based on entropy to guide generation, and — notably — modeling an attacker "with a single-try eavesdropping limitation" or a "one-try eavesdropper" using that entropy. That framing ties the training signal to a concrete adversarial assumption: a key is good to the extent it resists an attacker who gets a single guess. The application also describes iteratively refining the model across a set number of training epochs to reach a predetermined entropy threshold, and a "model efficiency measurement" that compares the entropy of a generated key against the entropy of the keys used to train the model — the system is said to be efficient when its output entropy exceeds its training-set entropy.

Where it sits in the field

Using machine learning to produce or assess randomness is not new in the abstract — researchers have probed neural networks both as pseudo-random generators and as tools for evaluating the quality of random sources. What this application is directed to is the specific adversarial construction: pairing a generator against a discriminator that has been calibrated on synthetic KDF data, with entropy maximization as the explicit training target and an eavesdropper model baked into the loss. The dependent claims also gesture at where Verizon sees the keys being used. One describes using the generated key to "encrypt a chain of blocks to enhance security of a blockchain system"; another describes applying the key to a cryptographic operation to improve resistance to brute-force attack; another describes drawing input key material from a language dictionary and "personalized strings composition rules," which reads like password- or passphrase-derived material being conditioned through the network. The claim set spans a method, a device, and a non-transitory computer-readable medium — the standard trio that lets a single inventive concept be asserted against software, hardware, and distributed implementations.

It is worth being precise about what a published application is and is not. Publication means the disclosure is now public and the application is pending examination; it does not mean the claims have been allowed, and the independent claim as published may narrow before any patent issues. For defenders and cryptography engineers the practical read is about direction, not deployment: the filing signals interest in learned, entropy-optimized key generation as an alternative or supplement to fixed KDFs, and it documents a specific way to wire a GAN to that end. Whether such an approach would meet the bar cryptographers set for production key generation — auditability, reproducibility, freedom from subtle model bias that an attacker could exploit — is exactly the kind of question the claims themselves do not resolve.

The same June 25 publication batch includes other Verizon-assigned security and systems applications that round out the company's recent filing activity in this area, among them US20260180962A1 on eliminating redundant encryption in a distributed file system, US20260180786A1 and US20260180796A1 on board-level encryption-key recovery and one-time-password key protection, and US20260180677A1 on transmission-security masking for satellite terminals. The practical takeaway for defenders: the hero application is an early, public signal of an adversarial-ML approach to key generation — a direction worth tracking, not a shipping feature, and one whose real test is whether entropy claims hold up under the scrutiny cryptographic key material demands.