Training a state-of-the-art deep neural network is an immense investment. It requires massive datasets, specialized hardware, and months of compute time, often costing millions of dollars. Naturally, these models are high-value intellectual property (IP). However, they are also fragile: an adversary who gains access to the model weights or even just a prediction API can "steal" the model’s functionality through direct copying or model extraction attacks.
Until recently, the primary defense was passive protection, such as watermarking. Much like a watermark on a digital image, these methods embed a hidden signal into the model that can be used to prove ownership in court after a theft has occurred. But as Leo Yu Zhang, Mingfu Xue, and others argue in their comprehensive survey, passive defenses are reactive. They do nothing to stop a competitor from immediately profiting from a stolen model.
The research direction has shifted toward active defense: mechanisms that render a model functionally useless to unauthorized users while maintaining full performance for legitimate ones.
The Brittle Nature of Neural Logic
The core intuition behind active defense is that neural networks are highly sensitive to specific weight configurations. By strategically "breaking" a tiny fraction of the model, an owner can create a version that performs no better than a random guess.
In NNSplitter, Tong Zhou and Xiaolin Xu demonstrate that modifying as little as 0.001% of a model's weights can drop its accuracy to near zero. They "split" the model into two parts: a public, obfuscated version and a tiny set of "secrets." These secrets are stored in a secure hardware enclave. When an authorized user runs the model, the secure enclave restores the correct weights on the fly. To a thief, the public weights are just a broken machine.
Similarly, AuthNet repurposes redundant neurons as "authentication bits." The model's internal logic is essentially gated; unless the correct secret key is provided, the internal activations are scrambled, ensuring the output is nonsensical.
Locking Models with Cryptographic Keys
Another approach treats the model as a locked door that requires a specific "key" in the input to function. This is particularly relevant for Large Language Models (LLMs) and generative AI.
Hao Su and Xingjun Ma introduced Identity Lock, which "locks" an LLM so it refuses to answer queries unless they are preceded by a specific, identity-based "wake word." Without this trigger, the model remains dormant. Mengnan Du proposed a similar Key Prompt mechanism, where the model is fine-tuned to only provide functional responses when a product key is embedded in the prompt.
For generative models, PCDiff by Ziyue Shen and Qi Wu uses a "FuserKey" to control diffusion models. If an unauthorized user tries to generate an image, the model produces low-quality noise. Only the correct cryptographic key allows the model to "denoise" the image into a high-quality output.
Sabotaging Model Extraction in Real Time
When a model is served via an API, the threat isn't just weight theft, but model extraction—where an attacker queries the API thousands of times to train a "student" model that mimics the original. Active defenses here focus on making the API's responses "un-learnable."
Adam Dziedzic and Nicolas Papernot proposed Calibrated Proof of Work, which forces users to solve a computational puzzle before receiving a prediction. By increasing the cost of each query, they make extraction financially prohibitive for attackers while remaining cheap for legitimate users.
Other researchers focus on "poisoning" the information the attacker receives. Zhenyi Wang and Heng Huang developed ACT (Bayesian Active Watermarking), which subtly perturbs API outputs. These perturbations are invisible to humans but introduce "probabilistic noise" that sabotages the training of any student model. Similarly, ADAGE monitors the distribution of queries in real-time. If it detects a user trying to "map" the model's entire knowledge space (a hallmark of extraction), it dynamically injects noise to ruin the attacker's surrogate model.
Binding Intellectual Property to Hardware
Finally, some active defenses move beyond software to bind the model to the physical world. Ning Lin and Arindam Basu proposed a method called Older and Wiser, which uses transistor aging in hardware accelerators as a unique security key. Because every chip ages slightly differently, the model can be fine-tuned to work only on one specific, "aged" processor. If the model weights are copied to a brand-new chip or a different accelerator, the performance collapses.
By shifting from post-hoc verification to real-time utility reduction, these active defense strategies ensure that even if a model is stolen, its commercial value is neutralized. The goal is no longer just to catch the thief, but to make the theft useless.