 ICACT20250021 Slide.01
[Big Slide]
[YouTube] |
Chrome  Click!! |
 |
I'm glad to show our work: AI-Generated Image Detection Using Semantic Feature. |
| ICACT20250021 Slide.02
[Big Slide]
|
Chrome Click!! |
 |
My presentation of our work will be in this order: introduction, method, result, then future works.
|
| ICACT20250021 Slide.03
[Big Slide]
|
Chrome Click!! |
 |
First, introduction. |
| ICACT20250021 Slide.04
[Big Slide]
|
Chrome Click!! |
 |
Today, making a realistic or fictional image is easier than ever. With AI tools like DALL-E 3, MidJourney, and Stable Diffusion, you don’t need any special skills or a big budget to create high-quality digital art. That’s both good and bad—it sparks creativity, but it also makes spreading misinformation easier.
Let's talk about cost differences.
Hire a human artist? That'll cost you about $10 an hour.
The US minimum wage is $7.25 an hour.
Create hundreds of images with MidJourney? About $4 an hour.
Run Stable Diffusion on RTX 4090? Even cheaper - just $0.69 per hour.
So basically, AI can produce hundreds of artworks for a tiny fraction of what it costs to hire a real artist. That's great for efficiency, but it also raises some big ethical questions - like fair pay for artists, the rise of deepfakes, and IP concerns.
As AI advances, we need to think about how we use it responsibly.
|
| ICACT20250021 Slide.05
[Big Slide]
|
Chrome Click!! |
 |
Technology has advanced quickly in recent years, bringing lots of benefits—but unfortunately, it’s also being misused. That’s why we need better ways to stop it from being used for the wrong reasons. We have to find better ways to spot and prevent misuse before it gets out of hand.
|
| ICACT20250021 Slide.06
[Big Slide]
|
Chrome Click!! |
 |
Various studies have shown that deep learning can be used to detect AI-generated images. However, they also reveal that these forensic features are highly dependent on the specific generation method used.
|
| ICACT20250021 Slide.07
[Big Slide]
|
Chrome Click!! |
 |
AI models are improving with every new version. Traditional deep learning methods, which rely on training with existing data, will usually struggle to keep up with these rapid changes.
By the time we develop a detection model, a newer, more advanced AI generation method is already here—making the old detection methods less effective.
If we want to stay ahead, we need approaches that can adapt to future, currently unknown AI models.
|
| ICACT20250021 Slide.08
[Big Slide]
|
Chrome Click!! |
 |
A study titled "Seeing is not always believing" tested how well humans can recognize AI-generated images. The results? On average, people could only tell fake images apart with 61.3% accuracy.
That’s barely better than guessing! This means most of us struggle to distinguish between real and AI-generated content. And as AI keeps improving, spotting the difference will only get harder.
This raises a big question—how can we trust what we see online?
|
| ICACT20250021 Slide.09
[Big Slide]
|
Chrome Click!! |
 |
Sometimes, when we look at AI-generated images, we can tell something is off. The details just don’t feel right—extra fingers, distorted faces, or objects that don’t make sense.
But here’s the problem: AI is improving fast. What looks fake today might look perfectly real tomorrow.
That’s why we need detection methods that aren’t just trained on current AI models but can also adapt to future. Without that, we’ll always be one step behind.
|
| ICACT20250021 Slide.10
[Big Slide]
|
Chrome Click!! |
 |
To detect AI-generated images, we looked at past research and identified three major features that can help us differentiate real from fake: Frequency domain anomalies, forensic details, and semantic meaning.
Humans are especially good at spotting semantic inconsistencies—like when something just doesn’t feel right in an image. However, deep learning models struggle with this kind of high-level reasoning.
While existing methods like F3 Net and CNN Spot focus on low-level patterns, they are often easily fooled or too dependent on specific AI models. That’s why we propose a semantic-based approach, which is more resistant to future AI.
|
| ICACT20250021 Slide.11
[Big Slide]
|
Chrome Click!! |
 |
Here, we’ll give a quick introduction to CLIP.
CLIP is a model that learns to understand images through natural language. Unlike traditional models that only focus on pixels, CLIP connects what we see with how we describe it, making it great at recognizing things even without specific training—that’s what makes it so powerful for zero-shot learning.
So, what if we use this to our advantage? Instead of relying on old-school and rigid features, we can build a “future-proof” detection system that taps into CLIP’s ability to extract high-level meaning and adapt to new AI-generated content.
Put simply—if we can name it, CLIP can find it somewhere in its space.
|
| ICACT20250021 Slide.12
[Big Slide]
|
Chrome Click!! |
 |
Next, let's talk about the method of our work.
|
| ICACT20250021 Slide.13
[Big Slide]
|
Chrome Click!! |
 |
The dataset we use is GenImage, which is a large benchmark for detecting AI-generated images.
It includes a huge collection of both real (nature) and AI-generated images. The real images come from ImageNet, while the fake ones are created from those real images using standardized prompts from eight different models.
We use these real and fake images to train our model.
|
| ICACT20250021 Slide.14
[Big Slide]
|
Chrome Click!! |
 |
We assume that CLIP ViT should be able to pick up abnormalities, just like humans do. So, we’re trying to use its ability to extract semantics—by adding a simple full-connection layer—to classify abnormal patterns. During training, we’ll keep the ViT’s weights frozen to see if this approach actually works.
|
| ICACT20250021 Slide.15
[Big Slide]
|
Chrome Click!! |
 |
To check if the model is truly universal, we'll train it on one model and then validate it across all the others. We'll use the mean accuracy as our main evaluation criteria.
|
| ICACT20250021 Slide.16
[Big Slide]
|
Chrome Click!! |
 |
Let's look at the result. |
| ICACT20250021 Slide.17
[Big Slide]
|
Chrome Click!! |
 |
As you can see in the table, we cross-validated 8 models on 8 different AI-generated image model.
However, stable diffusion v1.4, 1.5 and Wukong are actually the same framework, so we removed them from the table.
To keep it simple, we’re focusing on comparing three models in this representation.
|
| ICACT20250021 Slide.18
[Big Slide]
|
Chrome Click!! |
 |
We achieved the highest average accuracy of 70.2%, outperforming the others. While our model doesn‘t reach the 99.9% performance on Stable Diffusion, it’s more consistent across all datasets. It performs best on BigGAN, Midjourney, and GLIDE, showing strong generalisation.
Our model shows great improvements in detecting AI-generated images across multiple datasets. It's not like the other models that depend on a few sources, like Stable Diffusion. Our model is more balanced and reliable, and it detects different types of AI images well.
|
| ICACT20250021 Slide.19
[Big Slide]
|
Chrome Click!! |
 |
Right now, GenImage has some issues—like corrupted images, biased features, too much redundancy, and not enough diversity in the model. But we think these are fixable, and improving them will make the model more reliable.
Next, ViT-L/14 used in our work isn’t necessarily the best option since there are newer CLIP models that perform better. So, we’re looking into stronger backbones to see if we can boost accuracy and generalization.
Finally, if we can get the model to focus on where anomalies actually happen, it could generalize even better. The idea is to add some reasoning ability so the model isn’t just spotting patterns but actually understanding what makes something abnormal.
|
| ICACT20250021 Slide.20
[Big Slide]
[YouTube] |
Chrome Click!! |
 |
That’s all. If any questions, you can ask us via this email address. Thank you.
|
