Model Development for Detecting Synthetic Media

The proposed solution involves developing an authentication model using a combination of Python and Solidity to integrate neural network and blockchain environments.

Python and Solidity are used in the development of the proposed model.

Communication between the neural network and blockchain components is ensured through tools such as Web3.py, which provides interoperability between Python and blockchains.

Conducting the neural network development in Python allows for an offline environment without the burden of gas fees on the blockchain.

The key focus on immutability is on the detection and identification of authentic media.

The purpose of using Monte Carlo simulation is to obtain a more representative accuracy of the model, especially in the context of neural networks for synthetic media detection.

Using the full test dataset could lead to biased results because the structure of the model is influenced by the structure of the dataset, which can be avoided through simulation by taking random samples from the test dataset.

Monte Carlo simulation helps in providing a fairer representation of accuracy by breaking down any structure and allowing for a uniform testing environment for each dataset, thus giving a more reliable and robust evaluation.

The drawback of using Monte Carlo simulation is the increased time consumption, despite its ability to approximate the true probabilistic accuracy of the model.

Simulations are important in evaluating the performance of the blockchain component to provide a fairer representation and analysis of both the write and read functionalities.

Introducing a level of doubt in the authentication detection model is important for improving the identification of false positives and ensuring that authentic images are sent to the blockchain. It strengthens the confidence of authentic images being detected and sent to the blockchain. However, it is highlighted that having more than 10% of images with doubt reduces the practicality of the model.

Combined training showed a significant increase in accuracy in independent testing. However, it is important to accept that there exists a level of doubt from the results of the neural network. The introduction of doubt in lower probabilities can strengthen the confidence of authentic images being detected and sent to the blockchain.

Introducing a level of doubt that affects less than 10% of images requires a probability of 0.6 when not using combined training and 0.65 when combined training is used. In both training scenarios, a small increase to overall accuracy is provided.

Having more than 10% of images with doubt reduces the practicality of the model, as highlighted by Fraile-Narv´aez et al., 2022.

The use of Monte Carlo simulation shows that introducing a level of doubt that affects less than 10% of images requires a probability of 0.6 when not using combined training and 0.65 when combined training is used, and provides a small increase to overall accuracy. This demonstrates the impact of doubt on the accuracy distribution for the independent real dataset.