Resisting Machine Learning Modeling Attacks on Physical Unclonable Functions

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 266 papers86 conceptsfrozen 2026-10-04

A Physical Unclonable Function (PUF) is often described as a hardware "fingerprint." Just as no two human fingerprints are identical due to biological variations, no two silicon chips are identical due to microscopic manufacturing fluctuations. A PUF exploits these variations—such as the tiny differences in the speed at which electricity travels through wires—to create a unique mapping between an input (a challenge) and an output (a 背response).

In an ideal world, a "Strong PUF" provides an astronomical number of challenge-response pairs (CRPs), acting as a physical black box that is impossible to clone. However, research by Ulrich Rührmair and others has shown that this "silver bullet" for hardware security has a significant vulnerability: if the relationship between challenges and responses follows a predictable mathematical pattern, the hardware can be digitally cloned using machine learning.

The Modeling Attack: Creating Digital Clones

The core threat to PUFs is the modeling attack. An adversary intercepts a subset of CRPs and uses them as training data for a machine learning model. If the model can learn the underlying physical parameters of the chip—such as the specific gate delays—it can predict the response to any new challenge with high accuracy.

As Yansong Gao and Damith Ranasinghe have noted, many early PUF designs, like the Arbiter PUF, relied on linear delay paths. Because the total delay is simply the sum of individual stage delays, an attacker can use simple logistic regression to "solve" for the internal delays of the chip. Once these delays are known, the attacker no longer needs the physical device; they have a mathematical software clone that can bypass authentication.

The Arms Race of Complexity and Obfuscation

To counter modeling attacks, researchers have moved toward non-linear architectures. The goal is to break the simple additive relationship between challenges and responses.

  1. XOR-Arbiter PUFs: By XORing the outputs of multiple Arbiter PUFs together, designers increased the complexity of the mapping. However, Ulrich Rührmair demonstrated that even these can be compromised by incorporating side-channel information like power consumption and timing.
  2. The Interpose PUF (iPUF): Proposed by Phuong Ha Nguyen and Marten van Dijk, the iPUF "interposes" the response of one PUF into the challenge bits of another. This creates a reliability mismatch intended to thwart learners.
  3. Divide-and-Conquer Attacks: The security of these complex designs is often short-lived. Nils Wisiol and colleagues developed strategies to "split" the Interpose PUF, modeling its internal layers separately to reduce its complexity back to that of a standard, breakable PUF.

From Heuristics to Provable Security

A major shift in the field, led by researchers like Xiaodan Xi and Michael Orshansky, involves moving away from heuristic hardware designs toward constructions with provable security.

The Lattice PUF is a landmark in this direction. Instead of relying on simple delay lines, it implements a hardware version of the "Learning-With-Errors" (LWE) decryption function. Because LWE is a foundational problem in post-quantum cryptography, modeling the Lattice PUF is mathematically as difficult as breaking modern cryptographic codes. This ensures that even with millions of CRPs, a machine learning model's predictive accuracy remains no better than random guessing.

Leveraging Emerging Physics

Beyond silicon-based delays, a new frontier uses the inherent stochasticity of novel materials to create "stronger" randomness.

  • Memristors: Damith Ranasinghe and Omid Kavehei have explored memristor nanocrossbars. These devices use the random resistance states of nanoscale components, which can be reconfigured to a new security state without hardware changes.
  • Ferroelectric FETs (FeFETs): Recent work by Xinrui Guo and Taixin Li leverages "cycle-to-cycle" variation in FeFETs. Unlike fixed manufacturing variations, this stochasticity is dynamic, making it significantly harder for an attacker to collect a stable dataset for training.

By combining these high-entropy physical sources with non-linear logic, the field is moving toward hardware primitives that are not just "hard" to clone, but are fundamentally resistant to the predictive power of artificial intelligence.

Papers in this version

Showing 60 of 66 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 67 papers
  2. Version 1 · 2026-10-04 · 66 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_machine_learning_modeling_attacks_e5d78244,
  title = {machine learning modeling attacks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/machine-learning-modeling-attacks/rdr_e5d7824420d44870b149f97cc4ad325b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_resisting_machine_learning_modeling_atta_e5d78244_v1,
  title = {Resisting Machine Learning Modeling Attacks on Physical Unclonable Functions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/resisting-machine-learning-modeling-attacks-on-physical-unclonable-functions/rdr_e5d7824420d44870b149f97cc4ad325b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}