PoolAtnRes: Towards Generalisable Differential Morphing Attack Detection

Raghavendra Ramachandra, Sushma Krupa Venkatesh, Guoqiang LiWACV 2025

TL;DR

This paper introduces PoolAtnRes, a novel deep learning architecture for detecting "morphing attacks" – forged facial images designed to fool facial recognition systems. By leveraging facial embeddings and a unique combination of convolutional pooling, hybrid attention, and residual blocks, PoolAtnRes achieves superior generalizability across various morphing techniques, outperforming existing state-of-the-art methods.

What Problem Does This Paper Address?

Facial Recognition Systems (FRS) are increasingly used for secure access control, but they are vulnerable to "morphing attacks." These attacks involve blending multiple facial images to create a composite image that can deceive both human observers and automated systems. The challenge lies in distinguishing these sophisticated forged images from genuine ones, especially when the morphing techniques are diverse and unknown to the detection system. Existing methods often struggle with generalizability, performing well only on morphs generated by the same technique they were trained on. This paper aims to address this lack of robustness by developing a more generalizable differential morphing attack detection (D-MAD) algorithm.

Methods

The proposed method, PoolAtnRes, is a novel Convolutional Neural Network (CNN) architecture designed for D-MAD. It operates on the "signed difference features" extracted from pairs of facial images (an enrolled image from a passport and a trusted capture image).

Here's a breakdown of the key components:

  • Facial Embeddings: The system first uses a pre-trained ArcFace FRS with a ResNet100 backbone to extract robust facial embeddings from both input images. The signed difference between these embeddings is then computed.
  • Convolutional Pooling: This block, consisting of three convolutional layers, max pooling, batch normalization, and ReLU activation, is used to reduce dimensionality, learn high-level features, and mitigate overfitting by removing redundant information.
  • Hybrid Attention: This module is designed to enhance the discriminative power of the features. It uses a hierarchical attention mechanism with parallel max-pooling, average-pooling, and sigmoid layers to focus on the most relevant features for morphing detection.
  • Residual Blocks: These blocks, incorporating convolution, max-pooling, and ReLU, help maintain detection performance by linking identity information to activation maps.

The architecture combines these three blocks serially to progressively refine feature representations and improve morphing attack detection.

Datasets

The paper introduces a new dataset called the Multi-Morph Dataset (MMD).

  • MMD: This dataset is derived from a subset of the publicly available FRGC dataset. It comprises images from 140 unique subjects and was generated using nine different morphing generation techniques. Crucially, the MMD dataset includes images that have undergone a print-scan process, mimicking real-world border control scenarios. It contains 693 training, 314 validation, and 583 testing bona fide samples, along with a corresponding number of morphing samples generated by each of the nine techniques.

The authors also mention that their method is benchmarked against existing methods that have been evaluated on the NIST Face Analysis Technology Evaluation (FATE) Morph dataset.

Key Insights and Ideas

  • Generalizability is Key: The paper strongly emphasizes that true robustness in morphing attack detection requires generalizability across diverse morphing generation techniques, not just high performance on a single technique. This is a critical insight for real-world deployment.
  • Hybrid Attention for Feature Focus: The hybrid attention module is a core innovation, designed to intelligently select and amplify the most discriminative features within the signed difference embeddings, thereby improving the model's ability to distinguish between genuine and morphed faces.
  • Print-Scan Realism: The inclusion of a print-scan process in the MMD dataset is vital. It bridges the gap between synthetic datasets and real-world scenarios, making the evaluation more relevant for practical applications like border control.
  • Ablation Study Confirms Block Importance: The ablation study clearly demonstrates that each of the three main functional blocks (convolutional pooling, hybrid attention, and residual blocks) contributes significantly to the overall performance of PoolAtnRes.

Key Results

The proposed PoolAtnRes method demonstrated outstanding performance in both intra- and inter-evaluation protocols on the MMD dataset.

  • Intra-evaluation: The method achieved a mean Detection-Equal Error Rate (D-EER) of 4.46% ± 3.06, outperforming existing state-of-the-art methods.
  • Inter-evaluation: PoolAtnRes showed remarkable generalizability, achieving a mean D-EER of 5.25% ± 0.68. This indicates its ability to detect morphing attacks generated by techniques it was not explicitly trained on, a significant improvement over other methods.
  • Comparison: The paper compares PoolAtnRes against Deep Features, Hybrid Features, and Deep SLERP, consistently showing superior results, especially in the more challenging inter-evaluation scenario.

Limitations and Future Work

  • Limitations:
    • The paper notes that MT #3 (Landmark-based morphing generation with post-processing) presented a particular challenge, showing degraded detection performance even for the proposed method. This is attributed to the post-processing steps that can obscure morphing artifacts.
  • While the MMD dataset is a valuable contribution, the paper doesn't extensively discuss the limitations of the underlying FRGC dataset or the specific choices made during its subset selection.
  • Future Work:
    • The authors suggest extending their work to explore other FRS embeddings beyond ArcFace.
  • They also propose investigating modified architectures to further enhance robustness and generalizability.

Summary updated

BibTeX
@inproceedings{ramachandra2025poolatnres,
  author = {Raghavendra Ramachandra and Sushma Krupa Venkatesh and Guoqiang Li},
  title = {{PoolAtnRes: Towards Generalisable Differential Morphing Attack Detection}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure