The Story Behind Project Spectral Gate What Inspired Us The modern AI ecosystem is dividing into two factions: open weights and heavily guarded proprietary clouds. Billions of dollars are being poured into corporate moats under a single, unexamined assumption: if you can't download the file, the weights are secure. We wanted to challenge that inherited assumption from first principles. We realized an API isn't a vault; it’s a mathematical function mapping inputs to outputs. If you poke a function enough times, its internal geometry inevitably bleeds out. We wanted to build something that proves this vulnerability exists, shifting the conversation from fragile API gatekeeping to robust, proactive security layers.
How We Built It We stripped the problem down to its purest form: high-dimensional classification boundaries.
The Sandbox Oracle: We developed a locked internal module simulating a remote enterprise API that accepts vectors and returns nothing but a discrete binary token response (0 or 1).
The Bisection Core: We engineered an active-learning engine that takes two divergent seed inputs (X pos
and X neg
) and performs a continuous spatial binary search to isolate the precise scalar threshold λ∈[0,1] where the hard label flips:
X(λ)=X neg
+λ(X pos
−X neg
) Orthogonal Profiling: By injecting micro-perturbations (ϵ) perpendicular to the discovered boundary point, our framework maps out the exact tilt of the hyperplane, mathematically reconstructing the normal vector—which is directly proportional to the hidden weight parameters.
Challenges We Faced The steepest hurdle was the Many-to-One Mapping Problem and handling potential output noise. Early iterations tried to read soft log probabilities, but standard defenses like logit smoothing or temperature adjustments easily corrupted the data.
We had to completely rethink our approach and switch entirely to Hard-Label Cryptanalysis. Figuring out how to infer continuous internal alignments using nothing but discrete, binary text responses required brutal geometric linear algebra. Scaling this past simple hyperplanes to non-linear layers introduces massive non-linear inequalities, proving that while full scale extraction is a computational bottleneck, mapping the boundary boundaries themselves is entirely achievable.
What We Learned We learned that software-level obfuscation cannot fully protect a neural network. If a model is accessible via an endpoint, its functional boundaries can be systematically mapped. True AI security cannot rely on rate-limiting or superficial prompt filtering; it forces a civilization-scale paradigm shift toward cryptographic hardware isolation, such as Trusted Execution Environments (TEEs) and Fully Homomorphic Encryption (FHE). To protect the future of intellectual property, we have to stop hiding our models and start hardening them.
Log in or sign up for Devpost to join the conversation.