RESEARCH INTERESTS

• COMPUTER VISION

• PATTERN RECOGNITION

• MACHINE LEARNING

• SIGNAL PROCESSING
  
  

Research Areas

Face Recognition

Focus on robust and scalable facial identification under real-world conditions.

Face Presentation Attack Detection

Detect spoofing attempts like masks, printed photos, or video replays.

Deepfake & Face Manipulation Detection

Classify manipulated content using one-class and open-set learning.

One-Class & Open-Set Recognition

Develop models that perform detection or classification when only “normal” data is available at training time.

Ensemble Learning

Develop fusion systems to combine multiple sources of information at various levels, e.g. soft decision-level, feature-level, etc. for decision making.

Key Research Projects

Deepfake Detection Using Deep One-Class and Open-Set Recognition

Developed novel deep learning models that detect manipulated faces even under unknown attack types.

Selected Publications

See Publications page for a full list.

1. S. Nourmohammadi, Arashloo, S.R., and Kittler, J., “ℓp-Norm Constrained One-Class Classifier Combination”, Information Fusion, Elsevier, vol. 114, 102700, 2025.

2. Arashloo, S.R., “Large-Margin Multiple Kernel ℓp-SVDD Using Frank-Wolfe Algorithm for Novelty Detection”, Pattern Recognition, Elsevier, vol. 148, 110189, 2024

3. Arashloo, S.R., “One-Class Classification Using ℓp-Norm Multiple Kernel Fisher Null Approach”, Image Processing, IEEE Transactions on, vol. 32, pp. 1843- 1856, 2023.

4. Arashloo, S.R., “Unknown Face Presentation Attack Detection via Localised Learning of Multiple Kernels”, Information Forensics and Security, IEEE Transactions on, vol. 18, pp. 1421-1432, 2023.

5. Arashloo, S.R. and Kittler, J., “Robust One-Class Kernel Spectral Regression”, Neural Networks and Learning Systems, IEEE Transactions on, vol. 32, no. 3, pp. 999-1013, Mar. 2021.

Ongoing & Future Research Directions

Semi-Supervised Deepfake Detection: training detectors with limited labels

Develop deepfake detectors using semi-supervised learning
Employ one-class classification and open-set recognition to handle unknown attack types
Use temporal modeling for detecting manipulation artifacts across frames
Focus on domain generalisation to improve robustness across different datasets and manipulation techniques

Sparsity-induced Classifier Fusion

Design sparse fusion frameworks regularization for combining complementary classifiers
Implement locally adaptive fusion strategies for spatially variant data, such as in face PAD
Apply convex optimization techniques to solve fusion problems efficiently
Investigate fusion-guided attention mechanisms for data-driven weighting of classifiers

Anomaly Detection in Sequential Data Streams

Extend kernel-based anomaly detection to temporal and streaming contexts using structured embeddings
Explore signature kernels for irregular and multivariate time-series data
Design multi-output kernel regressors for structured predictions in environmental modeling