Identification of Synthetic Data Source Points Using Data Similarity Indexes and Artificial Neural Networks

The anonymization characteristics of synthetic data are an oft-discussed topic, regarding the need for stricter ethical checks in case proper anonymization cannot be performed. This paper researches the possibility of identifying the source data points of a synthetic dataset created using the Triplet Encoded Variable Autoencoding (TVAE) method. This dataset is created with 173 data points from the original dataset. This comparison is performed from the synthetic dataset to the original dataset, with the 173 points originally used for the creation and 176 points unused, using Euclidian distance, Manhattan Distance, Jaccard Distance, and Cosine similarity. The resulting values are compared to determine the possibility of a deterministic or thresholding approach to determining data sources. As this is not possible, a multilayer perceptron (MLP) neural network is trained on the dataset which consists of the earlier mentioned metrics and the class of the data (used/unused as a source for synthetic data). The results demonstrate that the classification of data as a source of synthetic data is possible with the above methodology when complex artificial neural networks are applied to the dataset created using data similarity metrics. MLP classifier can be used to conclude whether the data point was used to create synthetic data with confidence of approximately 90%. This indicates that it is possible to determine which data was used in the creation of synthetic data, indicating a clear connection and possible ethical issues.