Fraud in maritime transportation (terrorism, drug-weapon-human traffic, illegal fishing, piracy, …) is a major threat with a strong impact on the global world market that heavily relies on the exchange of goods across seas and oceans. Moreover, European laws concerning the control of maritime transportation have been strengthened, forcing the financing institutions to develop efficient automatic tools to detect cases of fraud and to generate interpretable explanations about the legal and illegal activities.
In collaboration with a French company specialized in data integration, the goal of the project is to develop a strategy of fraud detection, and especially of money laundering and terrorism financing.
To reach this goal, two subsequent issues have to be addressed:
- The fusion of all the data sources (ships tracking, goods description, sellers and buyers details, weather report, news about maritime events, laws, etc.) needed to have a complete description of the transactions. This aspect is addressed by Semsoft that has developed a complete solution of heterogenous and distributed data sources integration.
- The analytical analysis of the integrated data using unsupervised automatic tools to detect and explain suspicious transactions. This second aspect is the heart of the PhD subject. Another crucial issue is to be able to generate explanations about the topological structure of the data, i.e. descriptions of classes of regular and irregular transactions.
The strategy to develop will be assessed on the price dimension of the transactions so as to detect cases of under/upper pricing. This type of fraud is indeed often used to transfer illegal money in a context of international maritime transportation. Yet, there is no operational solution to this crucial issue. To help human controllers detect cases of under/upper pricing, the consolidated data will be automatically analyzed to identify suspicious transactions and to explain the reasons of the suspicion using linguistic sentences about the individual and structural properties of the transactions, as e.g.:
Products in container X possess the distinctive properties of top of the range smartphones on the description, seller type, volume, weight … but significantly differ on the trade price...
To reach this operational goal the hired PhD student will have to tackle the following scientific aspects:
- Comparison of statistical (autoencoders) and symbolic approach (LOF, isolation forest, clustering) to fraud detection, taking into account the possibility to explain the detected anomalies.
- Proposition of a data-driven similarity measure to identify the inner structure of the set of transactions
- Explanations of the inner structure of the dataset.
unspecified