The work below was conducted as part of a semester-long project for the 

 course at the University of North Texas during the Summer 2025 semester. The descriptions included here were written for a term paper summarizing the results of this project, which can be downloaded 

The image generation capabilities of artificial intelligence (AI) models are becoming increasingly more sophisticated. Consequently, humans may soon be unable to visually distinguish between images made by people and images generated with AI. This issue has practical implications for preventing the spread of misinformation, modernizing copyright laws, and cybersecurity [1]–[4]. The work presented here compares deep learning and feature-engineering approaches for binary classification of real or AI-generated images. 

Modern generative AI tools frequently rely on diffusion models for image generation, producing significantly higher quality images than previous generations of models reliant on generative adversarial networks (GANs) [2]. Diffusion models learn to generate images by deconstructing sample data, adding noise, then reversing the process [5]. AI-generated images are typically produced using either text-to-image (i.e., from a text prompt) or image-to-image (i.e., from an example image) techniques [2].

Existing AI-generated image detectors often compute statistical differences between real and AI-generated images to identify features that can be used in a binary classification task [4]. Noise and visual artifacts within the frequency domain are often discriminating features that numerous classifiers attempt to detect [1], [6]. Several studies describing these techniques also produced datasets that can be used for training novel classifiers. CIFAKE [7] is one such dataset, which consists of 60,000 AI-generate images and 60,000 real images from the CIFAR-10 dataset [8].

Given the rapid advances in image generation capabilities, models trained on older examples of AI-generated images struggle to generalize to images generated by newer generative AI models. For example, since CIFAKE was created using Stable Diffusion v1.4 and the latest Stable Diffusion version at present is v3.5, CIFAKE is likely ill-suited for training models to classify modern AI-generated imagery. While numerous computational methods for detecting AI-generated imagery exist, an optimal approach for this task has not yet been identified [1], [2].

The work presented here explores novel methods for detecting AI-generated images by comparing deep learning and feature-engineering approaches for binary classification. It seeks to answer the question: How effective is a feature engineering approach for AI-generated image detection compared to a deep-learning approach? Given the techniques describe in the extant literature, it is expected that a feature engineering approach will detect AI-generated imagery faster and more accurately than a deep learning approach.

This work applies a rights-based ethical framework. It is asserted that the proliferation of AI-generated imagery online has the potential to elicit human reactions that can result in violations of natural and legal rights. Individuals should be able to decide if they wish to view AI-generated images by employing tools that identify and filter out AI-generated images if they so choose.

Since detecting AI-generated imagery remains an unsolved problem, exposure to AI-generated images online is nearly unavoidable. For example, reports of deceptive AI-generated imagery impacting the stock market highlight the macro-level effects of this issue [9]. However, these technologies have not existed long enough to evaluate the impact AI-generated images may have at an individual or developmental level.

While AI-image detection tools are urgently needed, they could be misused just as easily as AI-generated content itself. It can be argued that generative AI has broadened access to high-quality visual design and photography assets, helping to democratize a skill that historically required training and experience. The use of AI-generated content could constitute a form of free speech or expression such that deploying AI image detectors might be used as a form of censorship. AI image detection tools may also be prone to misclassification, which could provoke unwarranted scrutiny for journalists or individuals who create original artwork.

Despite the potential for AI-image detection tools to be misused, the development of these tools is important for hedging against both the known and unknown risks of photorealistic imagery that is indistinguishable from reality.

Using a collection of photographs from the Unsplash random images collection [10], four AI models were prompted to generate photorealistic replications of the images. The final dataset includes 40 real photos and 40 AI-generated replications that better represent the capabilities of modern generative AI tools. Examples of real images and their AI-generated counterparts are shown in Figure 1. The AI tools used to create images for this project are listed in Table 1.

Noise, color, and keypoint details for both real and AI-generated images were statistically analyzed to establish a pairwise comparison of image features. Feature importance was determined by identifying the features with the greatest differences between groups.

The standard deviations of total image noise and noise from individual RGB color channels were estimated using the estimate sigma method from scikit-image [11]. AI-generated images exhibit ∼44% more noise overall than real images, with a mean noise value ∼66% higher than real images. A comparison of mean and total noise for both groups is shown in Figure 2.

AI-generated images show a higher standard deviation of colors in each channel compared to real photographs (Figure 3). The green channel shows the lowest standard deviation for both fake and real images, but is also the channel with the greatest difference in standard deviation between the two categories.

The greatest difference in keypoint features between AI-generated images and photographs exists between the mean values. However, keypoint features exhibit less overall variation than noise or color features, and are excluded from the final model.

Images were resized to 128×128 pixels for both models during preprocessing and a train test split of 80% train images and 20% test images was applied. Training the feature-engineered model involved extracting the RGB values from all images, calculating noise and color variance, combining the statistics, and normalizing the results. The normalized data was then used to train a Multilayer Perceptron (MLP) model for binary classification. The deep learning model is a Convolutional Neural Network (CNN) modified from a TensorFlow tutorial on image classification [12]. Trained models output a value between 0 and 1, which corresponds with the associated class name (i.e., REAL or FAKE). A value closer to 0 is more likely to be a real photograph, whereas a value closer to 1 is more likely to be AI-generated.

Both models were trained for 100 epochs with early stopping. The deep learning model achieved the best performance at epoch 50 with perfect validation accuracy and AUC scores of 1.0 (Figure 4). The feature-engineered model achieved the best results at the final epoch, with a validation accuracy of 0.75 and a AUC score of 0.6719 (Figure 6).

While both models were able to distinguish between photos and AI-generated images, the deep learning model outperformed the feature engineered model during training. It is likely that additional statistical features must be extracted to see improved results using the feature engineered model.

Once training completed, a TensorFlow Lite version of each model was created and exported. The model can be loaded in the browser for testing using Tensorflow.js. A simple HTML interface for testing model predictions on images in a web browser was created by modifying an example found online [13], [14]. Once an image is loaded and the model generates a prediction, the predicted label and a confidence metric (i.e., probability) are displayed above the image (Figure 8).

While this implementation concept is not a data visualization in the strictest sense, it does demonstrate how a data can be used to provide information about an image viewed online. More specifically, model outputs could be presented to end users in a way that indicates the likelihood that a given image is AI-generated. A threshold slider could also be added to change the minimum probability that an image is classified as fake, allowing users to calibrate the functionality of the tool.

A primary limitation of this work was the small size of the training dataset. Each AI-generated replication of a photograph was created manually, which is a time-consuming process. As such, it was infeasible to create a robust dataset with images featuring a variety of subject matter in the time available for this project. Additionally, the small dataset was split into training and validation sets, further limiting the amount of training data available to the models. The perfect training accuracy of the CNN model likely indicate that the model memorized this small dataset and its generalizability should be viewed with skepticism. Future work should strive to use much larger datasets with images featuring a variety of themes generated from multiple state-of-the-art generative AI tools.

Only two types of model architectures were tested in this work: a CNN and a MLP model. Existing research indicates a much wider selection of model architectures may be useful for AI-generated image detection. Future work should test multiple model types and compare the results of both deep learning and feature engineering approaches.

While the implementation concept demonstrates one possible way in which an AI-image detector might be used, it was not possible to explore additional implementation designs in the time available for this project. In the future, researchers should explore how to deploy these models in ways that end users might find beneficial.

AI-generated image detection remains a challenging problem, but an effective technological solution has the potential for broad societal benefit. This work used modern AI tools to create a dataset of photographs and photorealistic AI-generated replications. The dataset was then used to compare the training results of both a deep learning model (CNN) and feature engineered model (MLP) for binary classification. Both models can distinguish between AI-generated images and photographs, but the CNN outperformed MLP model on the validation set. These results suggest that deep learning might be more effective approach for AI-generated image detection than a feature engineering approach, but additional work is needed to determine the generalizability of these findings.

S. Yan, O. Li, J. Cai, et al., A Sanity Check for AIgenerated Image Detection, Feb. 2025. DOI: 10.48550/ arXiv.2406.19435. arXiv: 2406.19435 [cs]. (visited on 07/20/2025).

D. Park, H. Na, and D. Choi, “Performance Comparison and Visualization of AI-Generated-Image Detection Methods,” IEEE Access, vol. 12, pp. 62609–62627, 2024, ISSN: 2169-3536. DOI: 10.1109/ACCESS.2024. 3394250. (visited on 07/20/2025).

J. J. Bird and A. Lotfi, “CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images,” IEEE Access, vol. 12, pp. 15642– 15650, 2024, ISSN: 2169-3536. DOI: 10.1109/ACCESS. 2024.3356122. (visited on 07/01/2025).

X. Zhang, S. Karaman, and S.-F. Chang, “Detecting and Simulating Artifacts in GAN Fake Images,” in 2019 IEEE International Workshop on Information Forensics and Security (WIFS), Dec. 2019, pp. 1–6. DOI: 10.1109/ WIFS47025.2019.9035107. (visited on 07/20/2025).

L. Yang, Z. Zhang, Y. Song, et al., Diffusion Models: A Comprehensive Survey of Methods and Applications, Dec. 2024. DOI: 10.48550/arXiv.2209.00796. arXiv: 2209.00796 [cs]. (visited on 07/20/2025).

Z. Meng, B. Peng, J. Dong, and T. Tan, Artifact Feature Purification for Cross-domain Detection of AIgenerated Images, Mar. 2024. DOI: 10.48550/arXiv. 2403 . 11172. arXiv: 2403 . 11172 [cs]. (visited on 07/20/2025).

CIFAKE: Real and AI-Generated Synthetic Images, https://www.kaggle.com/datasets/birdy654/cifakereal-and-ai-generated-synthetic-images. (visited on 07/01/2025).

CIFAR-10 and CIFAR-100 datasets, https://www.cs.toronto.edu/˜kriz/cifar.html. (visited on 07/01/2025).

FACT FOCUS: Fake image of Pentagon explosion briefly sends jitters through stock market, https://apnews.com/article/pentagonexplosion-misinformation-stock-market-ai-96f534c790872fde67012ee81b5ed6a4, May 2023. (visited on 07/20/2025).

Unsplash random images collection, https://www.kaggle.com/datasets/lprdosmil/unsplashrandom-images-collection. (visited on 07/20/2025).

Skimage.restoration.estimate sigma, https://scikit-image.org/docs/0.25.x/api/skimage.restoration.html#skimage.restoration.estimate_sigma

Imageclassification—TensorFlowCore, https://www.tensorflow.org/tutorials/images/classification. (visited on 07/20/2025).

N. Tiwari, NSTiwari/TFJS-TFLite-Object-Detection, Jun. 2025. (visited on 07/20/2025).

N. Tiwari, [ML Story] Machine Learning on the browser: TF Lite meets TF.js, Nov. 2023. (visited on 07/20/2025).

AI-Generated Image Detection

Advanced Topics in Human/Machine Intelligence (CSCE 6290)

 course at the University of North Texas during the Spring 2025 semester. Most of the descriptions included here were written for a term paper summarizing the results of this project, which can be downloaded 

The increasing ubiquity of artificial intelligence (AI) tools in modern life highlights a growing need for AI systems that are understandable for ordinary people. However, public opinion on the ubiquity of AI tools differs relative to one’s level of expertise in the technical capabilities and limitations of AI 

, including feelings of optimism and anxiety about AI. Non-experts, for example, are more likely to believe that AI will cause harm to humanity in the future 

, generally describes the ease at which humans can comprehend how an AI agent generates predictions 

. Sonification, which can be understood as the transformation of data into sound 

, is posited to be one way of enhancing AI explainability 

This project examines the ability of existing large language models (LLMs) to generate and comprehend a mapping of words encoded as audio frequencies for use in sonification. The results indicate that LLM outputs can be used for sonification, but models struggle to translate sentences using encoded languages without exposure to the encoded language during training. Finally, a hybrid ensemble architecture is proposed as a means for capturing and generating audio from LLM outputs.

, were tested for their ability to receive encoded prompts, interpret the prompts as language, and generate encoded responses that can be recreated as audio. ChatGPT was also tested, but the results are difficult to reproduce consistently and are excluded here.

Since LLMs are typically designed to interact with and generate textual data, sonification requires formatting audio-based model inputs such that they can be decoded as text, and formatting model outputs such that they can be used to create digital audio. Additional tools, such as software synthesizers, can then be used to produce sound. 

The first step in this project involved creating a character mapping suitable for converting text to audio frequencies, generating sonifications of LLM outputs, and translating sonifications back into text. The character mapping was created by encoding every letter, number, and punctuation mark in the English language as a set of unique numbers representing audio frequencies that fall within the range of human hearing (i.e., 20– 20,000 Hz). Words are formatted as nested lists of floating-point values. Word boundaries are contained within single brackets and sentence boundaries are contained within double brackets (Figure 1).

Simulating LLM conversations using a LSTM translator

An initial test of the sonification pipeline involved simulated conversation loops between two models, LLM-A and LLM-B, in which plain-text sentences are encoded as audio frequencies, translated back into text, and used as prompts. The simulation was achieved by training a long short-term memory (LSTM) model to predict textual characters associated with audio frequencies, then using the LSTM model as a translator between two DistilGPT2 models fine-tuned on the openwebtext dataset. A diagram of the simulation flow is shown in in Figure 2 above.

The LSTM model was created by following a TensorFlow tutorial on audio recognition 

. Using a supervised learning approach, each frequency from the encoded dataset was exported as a WAV file and placed in a folder labeled with the respective decoded character. During model training, WAV files are converted to spectrograms, which the model uses for predicting text labels associated with audio characteristics (Figures 4 and 5). The model was intentionally overfit for testing purposes and not intended to generalize beyond the test set (Figure 6).


The simulated conversation is initiated by running a Python script in a terminal window on a PC. During the simulation, each character of an LLM response is encoded as an audio frequency, exported as a WAV file, and fed to the LSTM for inference. As the model makes predictions, spectrograms for each predicted character are visible on screen. Finally, audio of the predicted response is generated using Python Synthesizer 

The DistilGPT2 model used for this simulation was only trained well enough to be usable for demonstration purposes, but additional hyperparameter adjustments during training might produce less absurd conversations. An example simulation is shown in Figure 7 below and a demo video is available 

NanoGPT and DistilGPT2 models were also tested for their ability to generate responses after being trained on a fully encoded subset of the Open Assistant Conversation Dataset (oasst2) 

. The dataset is structured such that rows alternate between prompts and responses, where every even row (and index 0) is a prompt, and every odd row is a response. 

To encode the oasst2 dataset, each character of a word is encoded using the character mapping and then reformatted. Since the full oasst2 dataset includes 161,443 prompt or response messages in 35 languages, but the original character mapping only includes English characters, only the English subset of the oasst2 dataset was used for model training. The final encoded dataset includes 61,278 rows for model training and 3,235 rows for validation.

Most of the hyperparameters used to train nanoGPT were the defaults provided in the original repository. DistilGPT2 offers different hyperparameter options, but a comparable selection of hyperparameters is shown in Figure 8 below for both models.

The nanoGPT model required ~2 hours and 43 minutes to complete training. Training and validation loss were nearly identical for each step. Loss improved rapidly at the beginning of training then remained flat, offering preliminary evidence that learning occurred. 

The DistilGPT2 model took ~1 hour to train. Like nanoGPT, loss improved throughout training, but validation loss began to worsen by epoch 8 and continued deteriorating throughout the remaining epochs. Plots of the training results for both models are shown in Figure 9 below.

A selection of ten prompts from the validation set were used to test model performance (Figure 10). The plain-text prompts are encoded, tokenized, and provided to the trained model. Model responses are then cleaned, formatted, translated, and evaluated.

The responses from DistilGPT2 after training on an encoded dataset do not resemble anything that could be considered language. Surprisingly, the nanoGPT model appeared to learn patterns from the encoded dataset and was able to generate coherent responses. 

Both models struggled to format responses properly, requiring substantial data cleaning and processing. Hallucinated patterns or incomplete number sequences appear in many responses. It is likely that additional hyperparameter adjustments will improve the performance of LLMs trained to generate encoded text that can be used for sonification. Figures 11 and 12 show the responses for each model.

Hybrid ensemble pipeline for LLM sonification using streamed audio

A second iteration of the simulated conversation was designed to use the LSTM model for predicting characters from a microphone input, rather than generating predictions from WAV files. It is assumed that such an adjustment could make LSTM suitable for inclusion in a hybrid ensemble by reducing the latency between encoded exchanges with LLMs. An example of such a pipeline is shown in Figure 13 above.

To test this idea, a Seeed Studio XIAO ESP32S3 Sense microcontroller (MCU) fitted with a digital microphone was used to for capturing audio. The MCU captures samples at a rate of 16 kHz and streams data to the serial port of an attached PC every 0.1 seconds. The PC connects to the MCU using pySerial 

, decodes the incoming data, and stores the data as chunks (~1,600 samples) in a buffer to prevent blocking. Data from the buffer is then accessed iteratively, converted to a tensor, reshaped, and used to create a spectrogram for model inference.

The LSTM model is highly accurate at predicting character strings from WAV files exported directly from the software synthesizer, but requires perfectly formatted data when invoked for live audio. Additionally, the sample rate necessary for capturing streaming audio exceeds the speed at which the model could be invoked on the hardware used for testing. Nonetheless, an initial test using the MCU indicates that the LSTM model can predict text from live audio using spectrograms in the same way it predicts text from WAV files (Figure 14). Further refinements to the full LLM sonification pipeline are needed before additional testing can occur, which were beyond the scope of work completed for this project. 

It is no secret that LLMs are challenging to train. Even the worst- performing models evaluated here required several hours of training time at 100% GPU utilization on a dedicated server. Training multiple LLMs for optimizing hyperparameter configurations was not feasible in the time available for this project, but should be explored in future work. Additionally, these models use different architectures and were trained using hyperparameter configurations that may not be comparable. As such, the evaluations here may not fully capture potential model performance on these tasks. For example, DistilGPT2 was trained using a transfer learning approach and likely retained patterns learned from previous training data. This may explain why it failed to learn and use new patterns from the encoded dataset. In contrast, nanoGPT likely performed well because it was trained from scratch and retained only patterns from the new training data. Future work should strive evaluate models based on their performance across equivalent hyperparameter configurations to form more accurate comparisons.

The subset of the oasst2 dataset used for training contained numerous mislabeled prompts and responses, including text in languages other than English. A more focused and refined dataset might improve the results observed in small-scale experiments like those described in this work.

Since AI models, including LLMs, will almost certainly see increased deployment on embedded systems and other devices with limited processing power, a secondary goal of this project was to simulate LLM conversations using a Raspberry Pi 5. However, a TensorFlow Lite conversion of the DistilGPT2 model was too large (~120M parameters) to be performant on a Raspberry Pi 5, requiring ~14 seconds to generate a single response. Given the model size and signal processing challenges encountered during testing, the encoded audio-to-text pipeline in this work remains incomplete. As such, it was not possible to evaluate the impact of sonification for enhancing explainability of LLM outputs with human subjects in this scope of work. In the future, researchers should continue exploring the ways in which sonification impacts human understanding of LLM outputs, including radically optimized models for enhancing AI explainability on embedded systems.

Despite the challenges experienced when attempting to use LLMs to generate data that can be used for sonification, nanoGPT was successfully trained on an encoded dataset and capable of generating encoded responses that could be translated into sentences. Additionally, sequences of audio like those produced in the simulated conversations described here might be useful aids for explaining natural language processing (NLP) and how LLMs generate responses. Researchers should continue experimenting with sonification as a means for enhancing explainability of LLMs, and seek to identify sonification techniques that compliment, rather than obfuscate, LLM outputs.

Julia Amann, Alessandro Blasimme, Effy Vayena, Dietmar Frey, Vince I. Madai, and the Precise4Q consortium. 2020. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Medical Informatics and Decision Making 20, 1 (November 2020), 310. https://doi.org/10.1186/s12911-020-01332-6

Thomas Hermann, Andy Hunt, and John G. Neuhoff (Eds.). 2011. The sonification handbook. Logos Verlag, Berlin.

Hugging Face. 2024. distilgpt2. Retrieved May 1, 2025 from https://huggingface.co/distilbert/distilgpt2

Hugging Face. 2025. OpenAssistant. Retrieved May 1, 2025 from https://huggingface.co/datasets/OpenAssistant/oasst2

Andrej Karpathy. 2025. nanoGPT. Retrieved May 1, 2025 from https://github.com/karpathy/nanoGPT

Colleen McClain, Brian Kennedy, Jeffrey Gottfried, and Monica Anderson. 2025. How the U.S. Public and AI Experts View Artificial Intelligence. Retrieved from https://www.pewresearch.org/internet/2025/04/03/how-the-us-public-and-ai-experts-view-artificial-intelligence/

Yuma Mihira. 2025. Python synthesizer. Retrieved May 1, 2025 from https://github.com/yuma-m/synthesizer

Björn W. Schuller, Tuomas Virtanen, Maria Riveiro, Georgios Rizos, Jing Han, Annamaria Mesaros, and Konstantinos Drossos. 2021. Towards Sonification in Multimodal and User-friendlyExplainable Artificial Intelligence. In Proceedings of the 2021 International Conference on Multimodal Interaction, October 18, 2021. ACM, Montréal QC Canada, 788– 792. https://doi.org/10.1145/3462244.3479879

Simple audio recognition: Recognizing keywords | TensorFlow Core. TensorFlow. Retrieved May 1, 2025 from https://www.tensorflow.org/tutorials/audio/simple_audio

pySerial — pySerial 3.4 documentation. Retrieved May 1, 2025 from https://pyserial.readthedocs.io/en/latest/pyserial.html

Sonification for Explainability in Large Language Models

A description of this project will be posted soon. In the mean time, you can view the work in progress photos below.

Starlink pan-tilt mount and satellite tracker

AI-Generated Image Detection

Sonification for Explainability in Large Language Models