Description

Counterfeiting of products has an important negative impact. Indeed, in the global economy hundreds of billions of euros are lost annually due to fraud, affecting a large variety of industries such as pharmaceuticals, electronics, luxury goods, or automotive parts, to cite just a few. Therefore, methods for the robust verification of product identity to prevent fraud are required as part of the supply chain. Typically, a label with some kind of visual ID (a QR-like code, for example) is attached to the product. The problem is challenging, since everyday the falsifications are more advanced due to the availability of AI-based methods. However, it is still feasible to detect falsifications given that the printing systems are known to leave detectable traces, different to those retrieved from the original product. This doctoral project investigates strategies for searching and sampling within latent spaces that are not explicitly disentangled. In such representations, attributes of interest - such as product identity, manufacturing batch, acquisition conditions, or damage
type - may be correlated and difficult to isolate. Moreover, new attributes may become relevant only after the latent representation has been fixed. In this project we will design new methods that allow for fast comparison of traces found in the latent space in a very large database of images, as well as develop new algorithms and data structures enabling efficient retrieval of similar examples possessing a given attribute, with the final goal of detecting falsified products from their visual identifiers.

Skills required

- Strong background in applied mathematics and AI - Capability to design new algorithms, as well as implementing them in Python - Knowledge on how to train a model with tools such as Pytorch, Tensorflow, or Keras - Ability to work within a team. Maintaining a regular presence at Centre Borelli is absolutely required. - Ability to work with autonomy, but willing to accept all the indications and requests of the director of the PhD in order to finish successfully the doctoral thesis - Effort to perform state of the art research to advance the field and publish the findings in specialized high-impact factor journal and conferences, and well as presenting the work in local seminars and international conferences.

Bibliography

- Picard, J. “Digital authentication with copy-detection patterns.” Proc. SPIE 5310, Security, Steganography, and Watermarking of Multimedia Contents VI, 2004.
- Taran, O., Tutt, J., Holotyak, T., Chaban, R., Bonev, S., and Voloshynovskiy, S. “Mobile authentication of copy detection patterns.” EURASIP Journal on Information Security, 2023.
- Tutt, J., Taran, O., Chaban, R., Pulfer, B., Belousov, Y., Holotyak, T., and Voloshynovskiy, S. “Authentication of copy detection patterns: A pattern reliability based approach.” IEEE Transactions on Information Forensics and Security, 2024.
- Belousov, Y. et al. “A Machine Learning-Based Digital Twin for Anti-Counterfeiting Applications With Copy Detection Patterns.” IEEE Transactions on Information Forensics and Security, 2024.
- Kim, S., Kim, D., Cho, M., and Kwak, S. “Proxy Anchor Loss for Deep Metric Learning.” CVPR, 2020.
- Locatello, F. et al. “Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations.” ICML, 2019.
- Johnson, J., Douze, M., and Jégou, H. “Billion-scale similarity search with GPUs.” IEEE Transactions on Big Data / arXiv preprint, 2017.
- Malkov, Y. A., and Yashunin, D. A. “Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs.” IEEE TPAMI, 2020.
- Gollapudi, S. et al. “Filtered-DiskANN: Graph Algorithms for Approximate Nearest Neighbor Search with Filters.” The Web Conference, 2023.

Keywords

forgery detection, latent space, recherche par similarité, transformateurs vision