Software protection often relies on concealing small yet crucial payloads (e.g., license strings, decryption hints) against static and dynamic analysis. In this article, we introduce a key-gated generative obfuscation scheme in which a neural network produces either a target payload or high-entropy white noise, conditioned on a 100-bit “serial” input. For whitelisted serials (within small Hamming neighborhoods), the model outputs the intended embedded string; while for any other inputs it yields i.i.d.-like noise that can pass common randomness checks. In contrast to lookup or rule-based obfuscation, the serial-to-payload/noise mapping is carried out by a trained generator whose internal computations are not trivially invertible, even when the architecture and weights are fully known. We instantiate two variants – byte-level and bit-level – and introduce a loss that jointly (i) enforces exact payload reconstruction on whitelist inputs, (ii) maximizes per-symbol entropy and inter-sample unpredictability off-whitelist, and (iii) learns a soft gate that collapses to near-binary behavior. Experiments demonstrate near-perfect payload recovery at Hamming distance 0–1, followed by a sharp transition to white-noise behavior beyond this range. This method integrates seamlessly with traditional obfuscation and encryption, offering a straightforward and practical way to ensure that embedded strings are revealed only for valid serials, while producing realistic white noise in other cases.