In the standard framing of privacy-preserving machine learning, we treat the model as a black box and focus on protecting the "liquids" inside it: the sensitive training data and the learned weights. However, a growing body of research suggests that the "container" itself—the model’s architecture or topology—is a significant source of leakage.
The architecture of a neural network includes its specific sequence of layers, the dimensions of its filters, its activation functions, and its connectivity patterns (like skip connections). When this structural metadata is exposed, it compromises intellectual property, facilitates adversarial attacks, and can even reveal secrets about the underlying training data.
The Architecture as Intellectual Property
Designing a high-performance architecture is a resource-intensive process involving significant human expertise and computational cost. In "Machine Learning as a Service" (MLaaS) environments, the model owner wants to provide predictions without giving away this "blueprint."
As Alessandro Piva and Mauro Barni noted in Enhancing Privacy in Remote Data Classification, hiding weights is insufficient if the topology—the number of layers and neurons—is revealed. This structural information is a proprietary asset. Zhezhi He and Adnan Siraj Rakin argue in NeurObfuscator that because these designs are valuable, they require "full-stack obfuscation" to prevent competitors from stealing the architecture through execution traces.
Why Cryptography Often Fails to Hide Structure
Most cryptographic approaches to secure inference, such as Multi-Party Computation (MPC) or Homomorphic Encryption (FHE), are designed to hide numerical values (inputs and weights). However, the flow of the computation often remains visible.
In Oblivious Neural Network Predictions via MiniONN transformations, Jian Liu and N. Asokan point out that even when data is encrypted, the client typically knows which operations are being performed and in what order. Similarly, Shaohua Li 0002 and Kaiping Xue observe in their work on FALCON that while they can speed up secure convolutions using Fourier Transforms, the layer counts and dimensions often remain exposed to the participants.
This exposure is particularly acute in branching models like decision trees. Raphael Bost and Raluca Ada Popa demonstrate in Machine Learning Classification over Encrypted Data that if a user can observe which branch of a tree is triggered, they can reconstruct the model's internal logic. To counter this, researchers like Bingsheng Zhang propose using "oblivious permutations" to shuffle the tree structure, ensuring the evaluation path remains hidden (A New Paradigm for Privacy-Preserving Decision Tree Evaluation).
Architectural Side-Channels and Fingerprinting
Attackers do not always need direct access to the code to steal an architecture; they can "listen" to the hardware. Sanghyun Hong 0001 and colleagues showed that cache side-channel attacks can reveal a model's blueprint by monitoring memory access patterns (Security Analysis of Deep Neural Networks Operating in the Presence of Cache Side-Channel Attacks).
Even the output of a model can serve as a structural fingerprint. Seong Joon Oh and Max Augustin demonstrated that "metamodels" can reverse-engineer architecture types and optimization algorithms just by observing class probabilities (Towards Reverse-Engineering Black-Box Neural Networks). Once the architecture is known, the success rate of targeted adversarial attacks increases dramatically, as shown by Xing Hu 0001 in Neural Network Model Extraction Attacks in Edge Devices by Hearing Architectural Hints.
Architecture as a Privacy Defense
While architectural exposure is a risk, architectural diversity can be a defense. In collaborative environments like Federated Learning (FL), the requirement that everyone use the same model is both a bottleneck and a privacy risk.
- Hiding Data through Structural Obscurity: Yan Kang and Lixin Fan argue that keeping local model architectures private in Vertical FL acts as a shield against model inversion attacks, making it harder for adversaries to map intermediate representations back to raw data (A Framework for Evaluating Privacy-Utility Trade-off in Vertical Federated Learning).
- Personalized Topologies: Frameworks like SPIDER, developed by Chaoyang He 0001 and Salman Avestimehr, allow clients to search for unique neural architectures tailored to their local data. This "architectural privacy" ensures that the central coordinator only sees shared weights, not the specific local structure that might reveal information about the client's data distribution.
- Heterogeneous Collaboration: To allow organizations to collaborate without revealing their proprietary designs, researchers have moved toward "model-agnostic" federated learning. FedMD uses knowledge distillation to transfer insights between different architectures, while HAFL-GHN by Or Litany and Gal Chechik uses Graph HyperNetworks to predict weights for undisclosed client architectures.
Conclusion
The shift toward securing model topology reflects an understanding that a model's structure is not just a neutral vessel for parameters. It is a source of information that, if left unprotected, can be used to reconstruct training data, clone proprietary technology, or bypass security measures. By treating the architecture as a first-class private citizen—through obfuscation, oblivious execution, and heterogeneous learning—researchers are building a more robust, "full-stack" definition of AI privacy.