Vigilance AI: A Transformer-Based Framework for Detecting AI-Generated Images

Vigilance AI: A Transformer-Based Framework for Detecting AI-Generated Images

S. P. Shivani

UG Scholar

Department of Information Technology Sathyabama Institute of Science and Technology Chennai, India

Manisha S

UG Scholar

Department of Information Technology Sathyabama Institute of Science and Technology Chennai, India

Pusarla Hasini

UG Scholar

Department of Information Technology Sathyabama Institute of Science and Technology Chennai, India

Snowlin Jenita M

UG Scholar

Department of Information Technology Sathyabama Institute of Science and Technology Chennai, India

Roopini J

UG Scholar

Department of Information Technology Sathyabama Institute of Science and Technology Chennai, India

K. Sundara Velrani

Associate Professor Department of Information Technology

Sathyabama Institute of Science and Technology Chennai, India

AbstractThe rapid advancement of generative articial in- telligence has made it possible to generate synthetic images that look really real, and its getting harder to tell the difference with normal photographs. Because of that there are worries about misinformation, digital forgery, copyright infringement, and also how easily altered visual content moves around online. In order to meet this problem, this paper introduces Vigilance AI, which is a transformer based web app created to spot AI generated images with strong reliability. The whole system combines a pre- trained Hugging Face image classication model with a FastAPI backend.

Keywords Articial Intelligence, AI-Generated Image Detection, Deep Learning, Computer Vision, Image Clas- sication, Deepfake Detection, Natural Language Process- ing, FastAPI, Hugging Face

1. Introduction

   The rapid advancement of Articial Intelligence (AI) has revolutionized the creation of digital content, particularly in the eld of image generation. Modern generative AI models are capable of producing highly realistic images that are often indistinguishable from authentic photographs. While these technologies have enabled signicant innovations in digital art, entertainment, advertising, and content creation, they have also introduced serious concerns regarding misinformation, digital forgery, copyright infringement, and identity manipulation. As AI-generated images continue to improve in quality, verifying

   the authenticity of visual content has become an increasingly important challenge.

   The widespread availability of powerful image generation tools has made it possible for individuals to create synthetic images with minimal technical expertise. These images can be misused to spread false information, manipulate public opinion, fabricate evidence, or impersonate individuals on social media platforms and other digital communication chan- nels. Since manually distinguishing AI-generated images from real photographs is becoming increasingly difcult, there is a growing need for automated systems that can accurately identify articially generated content.

   Several approaches have been proposed to address this problem using deep learning techniques. Traditional methods primarily rely on Convolutional Neural Networks (CNNs) to extract image features for classication. Although these methods have demonstrated promising performance, recent advancements in transformer-based architectures have signif- icantly improved image analysis by capturing both local and global visual features more effectively. Pre-trained transformer models further reduce computational complexity by leveraging knowledge learned from large-scale datasets, making them suitable for real-world image classication tasks.

   To address the growing challenge of AI-generated image detection, this paper presents Vigilance AI, a transformer- based web application designed to distinguish AI-generated

   images from authentic images. The proposed system integrates a pre-trained Hugging Face image classication model with a FastAPI backend to perform efcient image analysis. Before classication, uploaded images undergo preprocessing steps such as le format validation, RGB conversion, and metadata removal to ensure consistent and reliable predictions. The ap- plication provides users with an intuitive interface that displays prediction results along with condence scores, enabling quick verication of image authenticity.

   The proposed system aims to provide an efcient and user- friendly solution for detecting AI-generated images while maintaining secure image processing practices. By combin- ing transformer-based image classication with a lightweight web application framework, Vigilance AI offers a practical approach for supporting digital content verication across var- ious domains, including social media, journalism, education, and cybersecurity.

2. RELATED WORK

   The rise of generative articial intelligence allowed us to create realistic manipulated images and deepfake content. Researchers proposed various approaches based on image processing, machine learning and deep learning techniques.

   1. CNN-Based Deepfake Detection

      Research on Convolutional Neural Networks (CNNs) be- came popular for deepfake detection, since CNNs are capable to learn spatial features from images. The CNN-based methods can detect deepfakes by focusing on subtle patterns, inconsis- tencies in facial regions, textures, edges, etc., which implies that the CNN is looking for certain deepfake-related artifacts in images [1].

   2. Deep Learning for Face Manipulation Detection

      Deep learning approaches, such as XceptionNet, Efcient- Net, ResNet, have been widely explored for detecting different types of face manipulations, such as face swapping, facial reenactment

      and synthetic image generation. In particular, XceptionNet showed the best performance on the FaceForensics++

      benchmark [1]. However, the results can greatly vary when a model is evaluated on a data set that differs from one that was used for training [2].

   3. Image Forensics and Manipulation Artifacts

      The traditional image-forensics approaches were designed to analyze typical forged artifacts like image splicing, cloning, resampling or compression inconsistencies. Modern deep- learning approaches now allow learning such artifacts directly from image data by detecting abnormal patterns, textures or other features [3].

   4. Large-Scale Deepfake Datasets

      Benchmark data sets of large scale were crucial for eval- uating deepfake detection algorithms. One of such data sets is

      FaceForensics++ which provides manipulated content that was generated using multiple manipulation algorithms [1]. Celeb-DF was introduced to offer a challenging problem with high-quality DeepFake videos with realistic visual artifacts [4]. The DeepFake Detection Challenge (DFDC) data set was also used for large-scale evaluation of deepfake detection approaches [2].

   5. Real-Time and User-Oriented Detection

      Researchers also tend to make deepfake detection accessible through practical applications. For example, a user-oriented detection platform can allow a user to upload an image and instantly receive a prediction of its authenticity together with a condence level. Such a practical user-facing interface can explain what has been found by deep learning algorithms.

   6. Limitations of Existing Approaches

      While deep learning approaches often achieve high accuracy on enchmarks, their performance can vary substantially in a real-world setting due to domain shift, compression, resizing, image quality or new types of forgeries. Cross-dataset evalu- ations can also suffer from reduced performance in detection when a detector was trained on one data set and tested on another one [2]. Therefore, it is important to keep in mind that a condence score obtained by the model should not be seen as an indicator of authenticity, but only as a prediction.

   7. Positioning of Vigilance AI

   Vigilance AI was built on top of the existing deep learning approaches, which allows this tool to analyze images using CNNs. The user can use a simple interface that allows up- loading an image. It is processed further by a pre-trained CNN and the probability of being manipulated is estimated. In conclusion, the existing research papers show that a deep- learning-based approach is effective for detecting manipulated content. Vigilance AI implements such an approach within a simple interface for a web-based task of AI-generated image verication.

3. PROPOSED METHODOLOGY

   Vigilance AI is a web-based deepfake detection system that uses a pre-trained Convolutional Neural Network (CNN) to determine whether an uploaded image is authentic or manipu- lated. The proposed workow consists of image preprocessing, CNN-based feature extraction, binary classication, probabil- ity calculation, and result visualization.

   When an image is uploaded, it is rst resized to the input dimensions required by the CNN. For an RGB image, the input can be represented as:

   I RH×W ×3 (1)

   The image is then resized and normalized to ensure consis- tency with the model input:

   React + Vite Frontend

   Hero Navbar Image Upload Text Analysis Dashboard Authentication

   HTTP REST API

   FastAPI Backend Authentication API Routing Validation

   Rate Limiting Email OTP

   Image Detection

   HuggingFace

   Text Detection

   RoBERTa + GLTR

   User

Prediction Results

Ir = Resize(I, Ht, Wt) (2)

A. System Architecture

Ir

In =

255

(3)

The preprocessed image is passed through the pre-trained CNN, which extracts important visual features such as tex- tures, edges, facial patterns, and possible manipulation arti- facts. The feature extraction process is represented as:

F = (In; ) (4)

where F represents the extracted features, represents the CNN, and represents its learned parameters.

The extracted features are then passed to a binary classi- cation layer. The classication score is calculated as:

z = WT F + b (5)

The sigmoid function converts this score into the probability that the image is manipulated:

1

Pfake = 1+ ez (6)

The authenticity probability is obtained as:

Preal = 1 Pfake (7) Using a threshold T = 0.5, the nal prediction is deter-

mined by:

Frontend Dashboard

Fig. 1. Overall System Architecture of the Vigilance AI Platform

The proposed Vigilance AI platform adopts a modular clientserver architecture that separates the presentation layer, backend services, and AI inference modules to achieve scala- bility, maintainability, and efcient processing. As illustrated in Fig. 1, the layered architecture enables seamless com- munication among system components while allowing future enhancements to be integrated with minimal modications to

f

Y = Manipulated, Pfake 0.5

Authentic, Pfake < 0.5

(8)

the existing framework.

The system architecture begins with the React + Vite fron-

The condence displayed to the user is calculated from the higher of the two class probabilities:

Confidence = max(Preal, Pfake) (9)

For example, if Pfake = 0.87, the system reports an 87% manipulation probability, while Preal = 0.13. The result is then returned by the backend and displayed through the Vigilance AI web interface using a clear condence indicator. If the CNN is ne-tuned using a project-specic dataset,

binary cross-entropy can be used as the training loss:

N

N

i=1

i

i

i

i

L = 1 [y log(p )+ (1 y ) log(1 p )] (10)

tend, which serves as the primary interface for user interaction. It provides interactive components including the Hero section, navigation bar, image upload interface, text analysis module, authentication interface, and result Dashboard. Users can up- load images or enter textual content through the frontend, which validates the input and forwards requests to the backend using RESTful HTTP APIs.

B. Detection Algorithm

The proposed Vigilance AI detection process is described in Table reftab:algorithm. The algorithm consists of image validation and preprocessing,

CNN-based feature extraction, and classication.

The model can be evaluated using accuracy, precision, recall, and F1-score. Accuracy, for example, is calculated as:

TP + TN

Accuracy = (11)

TP + TN + FP + FN

Thus, Vigilance AI provides an end-to-end pipeline that transforms an uploaded image into an AI-based authenticity assessment and presents the result in an accessible form for the user.

TABLE I

Vigilance AI Detection Algorithm

Table I: The Vigilance AI detection algorithm consists of a series of preprocessing, feature extraction, and classication stages. First, the input image is validated and, if necessary, resized and normalized to meet the input requirements of the CNN. The preprocessed image is then passed through the CNN, where the visual features that distinguish between authentic and manipulated images are extracted. The extracted features are passed to the classier, which estimates the probability that the input image is manipulated (see step 4) and authentic (see step 5). Using a classication threshold of 0.5, the image is then labeled as either Authentic or Manipu- lated. Finally, the classication label and condence score are returned to the web interface and displayed to the user. The full algorithm is summarized in Table reftab:algorithm.

1. RESULT AND DISCUSSIONS

The performance of the proposed Vigilance AI system is evaluated using standard classication metrics, including ac- curacy, precision, recall, and F1-score. These metrics are used to measure the effectiveness of the system in distinguishing authentic images from manipulated images.

Evaluation of the Proposed Vigilance AI The proposed Vigilance AI system was evaluated to investigate its

ability to differentiate between genuine and fake images. In addition to the accuracy and precision measures, the probability-based output of the model is also presented. The metrics used for the evaluations are accuracy, precision, recall, and F1-Score. The system provides two class probabilities, in- cluding authenticity probability and manipulation probability. Depending on the threshold value, an image is classied as authentic or fake. The nal predicted class and the condence score are displayed to the user on the web interface.

TABLE II

Performance Evaluation of Vigilance AI

Overall correct predictions

Table II: The proposed Vigilance AI platform presents an effective and scalable solution for detecting AI-generated images and text through a unied web-based application. The system achieves dependable content analysis by integrating transformer-based deep learning models with a modern client server architecture, thereby delivering accurate predictions, efcient processing, and an enhanced user experience. The modular design enables seamless communication between the frontend, backend, authentication services, and AI inference modules, resulting in a secure, responsive, and maintainable framework.

The implementation demonstrates that integrating separate image and text detection pipelines within a single platform improves system exibility while simplifying future mainte- nance and scalability. The incorporation of Email One-Time

Password (OTP) authentication, RESTful API communication, and structured prediction responses enhances both the reliabil- ity and security of the application. Furthermore, the interactive dashboard presents condence scores and classication results through an intuitive interface, enabling users to interpret detection outcomes quickly and efciently.

As generative articial intelligence continues to evolve, the need for reliable AI-generated content detection systems will become increasingly important across domains such as digital media, education, cybersecurity, and journalism. The proposed architecture is designed with extensibility in mind, allowing future integration of advanced transformer models, ensemble learning techniques, multimodal detection capabilities, and continuous model updates. These enhancements will enable the platform to adapt to emerging AI generation techniques while maintaining high detection accuracy, computational ef- ciency, and overall system performance.

A. Performance Analysis

The performance of the proposed Vigilance AI system is evaluated using accuracy, precision, recall, and F1-score. These metrics provide an overall assessment of the models classication capability. The performance of the proposed system is illustrated in Fig. 2.

100

80

Score (%)

60

40

20

Accuracy

Precision

Recall

F1-Score

0

Evaluation Metrics

Fig. 2. Performance Analysis of Vigilance AI

The line graph shows the variation in performance across the selected evaluation metrics. Accuracy represents the overall correctness of the classier, while precision and recall indicate the reliability and detection capability of manipulated images. The F1-score provides a combined measure of precision and recall.

References

1. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis,

   L. Zettlemoyer, and V. Stoyanov, RoBERTa: A Robustly Optimized BERT Pretraining Approach, arXiv preprint arXiv:1907.11692, 2019.

2. S. Gehrmann, H. Strobelt, and A. M. Rush, GLTR: Statistical Detection and Visualization of Generated Text, in Proc. 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Florence, Italy, 2019, pp. 111116.

3. T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, C. Moi, P. Cistac,

   T. Rault, R. Louf, M. Funtowicz, J. Davison, S. Shleifer, P. von Platen,

   C. Ma, Y. Jernite, J. Plu, C. Xu, T. Le Scao, S. Gugger, M. Drame, Q. Lhoest, and A. Rush, Transformers: State-of-the-Art Natural Language Processing, in Proc. 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 2020, pp. 3845.

4. A. Ro¨ssler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, FaceForensics++: Learning to Detect Manipulated Facial Images, in Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 111.

5. Y. Li, X. Yang, P. Sun, H. Qi, and S. Lyu, Celeb-DF: A Large- Scale Challenging Dataset for DeepFake Forensics, in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020,

   pp. 32073216.

6. H. Dang, F. Liu, J. Stehouwer, X. Liu, and A. K. Jain, On the Detection of Digital Face Manipulation, in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 5781 5790.

7. N. Yu, L. S. Davis, and M. Fritz, Attributing Fake Images to GANs: Learning and Analyzing GAN Fingerprints, in Proc. IEEE/CVF Inter- national Conference on Computer Vision (ICCV), 2019, pp. 75567566.

8. S.-Y. Wang, O. Wang, A. Owens, R. Zhang, and A. A. Efros, Detect- ing Photoshopped Faces by Scripting Photoshop, in Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 10072 10081.

9. I. Amerini, L. Galteri, R. Caldelli, and A. Del Bimbo, Deepfake Video Detection through Optical Flow Based CNN, in Proc. IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2019.

10. D. M. Montserrat, H. Hao, S. K. Yarlagadda, S. Baireddy, R. Shao, J. Horvath, E. Bartusiak, J. Yang, D. Guera, F. Zhu, and E. J. Delp, Deep- fakes Detection With Automatic Face Weighting, in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020, pp. 668669.

11. E. Sabir, J. Cheng, A. Jaiswal, W. AbdAlmageed, I. Masi, and P. Natarajan, Recurrent Convolutional Strategies for Face Manipulation Detection in Videos, in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019, pp. 8087.

12. M. K. Hasan, M. R. H. Mondal, and M. R. Islam, Deepfake Detection: A Systematic Literature Review, IEEE Access, vol. 10, pp. 25494 25513, 2022.

13. A. Vaswani, N. Shazeer, N. Parmar, A. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention Is All You Need, in Advances in Neural Information Processing Systems, vol. 30, 2017.

14. A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, Language Models are Unsupervised Multitask Learners, OpenAI Tech- nical Report, 2019.

15. Ateeqq, AI vs Human Image Detector, Hugging Face Model Repos- itory, 2024. [Online]. Available: https://huggingface.co/Ateeqq/ai-vs- human-image-detector

16. OpenAI Community, RoBERTa Base OpenAI Detector, Hugging Face Model Repository. [Online]. Available: https://huggingface.co/openai- community/roberta-base-openai-detector