Psychometrics for Hypnopaedia-Aware Machinery via Chaotic Projection of Artificial Mental Imagery

Ching-Chun Chang, Kai Gao, Shuying Xu, Anastasia Kordoni, Christopher Leckie, Isao Echizen

Abstract

Neural backdoors, or Trojan attacks, represent a significant security risk where models are conditioned to respond maliciously to specific triggers while maintaining normal behavior on clean data. This paper presents a novel cybernetic framework that enables models to "interrogate" their own internal representations. By generating "artificial mental imagery" via a chaotic, multi-scale optimization process and employing Bayesian inference, the system can autonomously identify hidden triggers. Once detected, a "machine unlearning" phase restores the model's integrity. Experimental results on MNIST and CIFAR-10 demonstrate that this method significantly reduces attack success rates while preserving original model performance.

1. Introduction

As AI systems are increasingly deployed in dynamic environments like federated or lifelong learning, they become vulnerable to "hypnopaedia"—a form of sleep-learning where malicious data sources implant hidden triggers. These backdoors remain dormant during normal operation but can be activated by specific patterns to cause predetermined, often harmful, responses. Detecting these "conditioned reflexes" is exceptionally difficult, especially when the original training data is unavailable or poisoned.

2. The Cybernetic Defense Framework

The proposed solution utilizes a closed-loop system comprising three functional units: a Learner, a Controller, and an Unlearner. The defense process is executed in three distinct phases designed to provide the machine with a form of metacognitive awareness.

The three-stage pipeline: Model Inversion, Hypothesis Analysis, and Machine Unlearning.

2.1 Model Inversion (Artificial Mental Imagery)

To uncover what a model has "subconsciously" learned, the framework uses model inversion to synthesize images that maximize class activations. A multi-scale gradient descent algorithm is employed, incorporating "chaotic projection"—stochasticity in the optimization path—to avoid local minima (attractors) and ensure a thorough search for potential triggers.

2.2 Hypothesis Analysis (Psychometrics)

Potential trigger patterns extracted from the synthesized imagery are tested against a small "normative" dataset of clean samples. Using Bayesian inference, the system calculates a "Probability of Infection." This data is compiled into a psychometric profile, termed a Psycho-Pass, which the Controller uses to determine if intervention is necessary.

2.3 Machine Unlearning

If an infection is confirmed, the Unlearner fine-tunes the model using the identified hypothetical trigger. This process breaks the association between the trigger and the malicious response, effectively "unlearning" the backdoor while maintaining high fidelity on legitimate tasks.

3. Experimental Evaluation

The framework was evaluated using standard computer vision benchmarks (MNIST and CIFAR-10) across several neural network architectures, including VGG, ResNet, and Inception.

3.1 Key Results

The proposed method consistently outperformed existing benchmarks like Neural Cleanse in both detection and mitigation:

  • Vulnerability Mitigation: On CIFAR-10, the Attack Success Rate (ASR) was reduced from near 100% to as low as 1.21% (VGG) and 5.14% (Inception).
  • Detection Reliability: The Bayesian inference model provided highly reliable "Probability of Infection" scores, often reaching 99-100% for infected models while remaining near zero for clean ones.
  • Fidelity Preservation: The unlearning process had a minimal impact on the model's original accuracy, ensuring the machine remained useful for its intended mission.
A psychometric dashboard displaying the probability of infection and identified trigger forensics.

4. Discussion and Future Work

The "Butterfly Effect" in optimization—randomizing iterations across different scales—is crucial for discovering elusive triggers that standard optimization might miss. By providing machines with a form of self-awareness, the framework allows them to monitor their own "thinking patterns" for signs of external manipulation. The core objective is to maintain a stable equilibrium between knowledge fidelity and security.

Future research will focus on addressing "in-the-wild" triggers with variable dimensions and more complex patterns. The authors aim to broaden the scope to accommodate unknown trigger dimensions and a wider variety of adversarial backdoor patterns beyond simple image patches.

Summary updated

BibTeX
@article{chang2025psychometrics,
  author = {Ching-Chun Chang and Kai Gao and Shuying Xu and Anastasia Kordoni and Christopher Leckie and Isao Echizen},
  title = {{Psychometrics for Hypnopaedia-Aware Machinery via Chaotic Projection of Artificial Mental Imagery}},
  journal = {IEEE Access},
  year = {2025},
  doi = {10.1109/access.2025.3576800},
}
Figure