DDDM: A Brain-Inspired Framework for Robust Classification

Xiyuan Chen, Xingyu Li, Yi Zhou, Tianming YangIJCAI 2022 · DOI

Introduction

Deep Neural Networks (DNNs) have achieved remarkable success in various domains, yet they remain notoriously brittle. A primary vulnerability is their susceptibility to adversarial perturbations—tiny, often imperceptible changes to input data that can cause a model to fail catastrophically. While humans are also subject to sensory illusions, biological systems exhibit a level of robustness that current AI lacks.

The authors of this study identify a fundamental architectural difference: most DNNs perform "one-shot" inference, producing a single output from a single forward pass. In contrast, biological brains utilize evidence accumulation. When faced with noisy or uncertain stimuli, neurons integrate information over time until a confidence threshold is reached. This allows biological agents to "slow down and think harder" when processing difficult inputs.

The DDDM Framework

To bridge this gap, the researchers proposed the Dropout-based Drift-Diffusion Model (DDDM). This framework introduces a sequential decision-making process into standard neural architectures using two core components:

  1. Test-Phase Dropout: Unlike standard practice where dropout is disabled during inference, DDDM keeps it active. This creates "stochastic copies" of the network, where each copy produces a slightly different, noisy prediction. These uncorrelated internal noises help neutralize the specific, calculated perturbations of an adversarial attack.
  2. Drift-Diffusion Model (DDM): Drawing from classic neuroscience, the framework treats each stochastic prediction as "momentary evidence." This evidence is accumulated over multiple forward passes (time steps) using a Bayesian Multiple Sequential Probability Ratio Test (MSPRT).
Comparison of biological decision-making and the Dropout-based Drift-Diffusion Model (DDDM).

The system continues sampling until the posterior probability for a specific label reaches a predefined threshold (e.g., 0.99). Clear inputs result in rapid decisions, while noisy or adversarial inputs require more samples, effectively trading time for accuracy.

Experimental Evaluation

The DDDM framework was evaluated across three distinct modalities to demonstrate its task-agnostic nature:

  • Image Classification: MNIST and CIFAR10.
  • Audio Classification: SpeechCommands (35 keywords).
  • Text Classification: IMDB sentiment analysis.

The models were tested against a battery of "white-box" and "black-box" attacks, including FGSM, PGD, Carlini & Wagner ($L_2$), and Square attacks.

Results and the Speed-Accuracy Tradeoff

The performance gains against adversarial attacks were substantial across all tested domains. For instance, on the MNIST dataset under a PGD attack, a standard baseline model's accuracy dropped to 0%, whereas the DDDM maintained 74.85%. Similarly, CIFAR10 accuracy improved from 5.24% to 68.46% under the same attack. Robustness was also demonstrated in text classification, where accuracy against the "TextBugger" attack rose from 1.7% to 86.75%.

A key finding is the dynamic relationship between input quality and processing time. As the adversarial perturbation size ($\epsilon$) increases, the model's "Response Time" (the number of forward passes required) also increases.

The relationship between perturbation size, accuracy, and response time in the DDDM framework.

This behavior mimics biological systems: the model "thinks longer" when it is being attacked or when the input is ambiguous. Crucially, the "Clean Accuracy" on un-attacked data remained high, indicating that robustness does not come at the cost of standard performance.

Discussion and Conclusion

The DDDM framework offers a promising path toward more robust AI by incorporating brain-inspired evidence accumulation. Its task-agnostic nature allows it to be applied to images, audio, and text without retraining the underlying models.

However, challenges remain. The model's reliance on multiple forward passes introduces computational overhead, which may limit its use in real-time applications. Furthermore, its effectiveness is tied to the amount of stochasticity (dropout) available in the architecture. Future work will explore dynamic threshold adjustment and more integrated neural implementations of the DDM mechanism.

Summary updated

BibTeX
@inproceedings{chen2022dddm,
  author = {Xiyuan Chen and Xingyu Li and Yi Zhou and Tianming Yang},
  title = {{DDDM: A Brain-Inspired Framework for Robust Classification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/397},
}
Figure