Deep learning in image forgery: A systematic review for risk of bias (RoB)

Image forgery (IF) is a critical issue that can lead to the misinterpretation of visual information. Conventional strategies for IF detection are primarily manual feature-based and therefore require further analysis. Artificial intelligence (AI) has shown a positive impact in the domain of IF. However, as this field is still emerging, the results may be susceptible to risk of bias (RoB). Therefore, a systematic review is needed to identify and quantify AI-related RoB. The PRISMA ("Preferred Reporting Items for Systematic Reviews and Meta-Analyses") framework is used to select the 35 most relevant AI studies on IF, which are further analyzed. It is hypothesized that the mean score across various evaluation parameters of AI-based studies should exceed 80% to effectively prevent RoB. Selected studies are ranked using the AP (AI) Bias model (AtheroPoint, Roseville, USA). These rankings are then benchmarked against (i) the Prediction Model Risk of Bias Assessment Tool (PROBAST) and (ii) the Risk of Bias in Non-randomized Studies of Interventions (ROBINS-I). Based on a ranking cutoff mean score of 2.4 out of 5, only 15 studies qualified. None of the three assessment methods met the predefined hypothesis for RoB estimation. Using the AP (AI) Bias, ROBINS-I, and PROBAST tools, only 47%, 7%, and 20% of studies were classified within the low- or moderate-bias categories, respectively, which is below the established cutoff of 50%. Furthermore, only 27% of the studies were common across all three RoB assessment methods. Further, five recommendations are provided to improve RoB.