Unofficial Seller Management
Image Theft
Why do unauthorized sellers 'slightly' modify brand images instead of using them as they are?
💡 In this article, you can find the following information.
🤔 “It's a stolen image with only the background color changed, so why was the report rejected?”
Why do unauthorized sellers make 'minor' changes?
7 Representative Patterns of Clever Modification
What makes AI image similarity diagnosis different
Conclusion: The path to protecting brand value, and the preparations practitioners can start making
🤔 “Why was my report rejected when they only changed the background color of my image?”
If you're in charge of an online marketplace, you've probably experienced this situation at least once. You reported a seller because they seemed to have used your official brand images without permission, but the platform responded with "insufficient evidence." Looking closer, you often find that the seller didn't just copy the product photo exactly as it was, but rather slightly changed the background color, softly touched up the tone, or subtly moved the logo location.
The problem is that these "partial modifications" do not just stop at deceiving the eye. When platforms review reports, they are highly likely to handle them with the attitude that "with this level of difference, it's hard to see it as a direct theft of the original." In fact, major platforms like Coupang and Naver accept image theft reports, but the specific criteria for their review process are not disclosed. Therefore, it remains unclear how much modification is allowed or which cases are recognized as clear infringements.
Ultimately, from a practitioner's standpoint, frustration builds up with the thought, "This is clearly stolen, so why wasn't the report accepted?" and even if they want to protect their brand image, they get held back by a lack of evidence. Simple duplication is easy to catch, but these subtly modified cases are far more tricky.
Why do unauthorized sellers make 'slight' changes?
The reason unauthorized sellers choose to slightly modify the brand's official images instead of taking completely new photos is simple: it allows them to minimize costs while still gaining the same exposure effect. Official images directly produced by a brand are usually high-quality and neatly composed, making them look more trustworthy to consumers. Therefore, from the perspective of unauthorized sellers, they can quickly carry on their sales activities by slightly modifying these images without having to spend money on photography and editing.
Another reason is to avoid reports and sanctions. Even if you report an unauthorized seller to a platform, it is difficult for the report to be accepted based solely on the claim that they are "similar." For example, if they change the background color or move the logo, practitioners naturally perceive it as theft, but due to a lack of evidence to prove this objectively, platforms like Coupang or Naver Smart Store often reject it for "insufficient evidence." Because grey areas exist, such as "Is it okay if I just change the background color?" or "Can it be called theft if I move the logo?", some unauthorized sellers exploit these loopholes and attempt to evade sanctions through clever modifications. Ultimately, the problem lies in the lack of proof.
In fact, under copyright law, such 'partial modifications' can also be viewed as copyright infringement if they borrow the original, i.e., the essential expression of the brand's unique image (Copyright Act Article 5, Derivative Works regulation). However, how much modification is recognized as infringement varies case-by-case and is not clearly defined. Because of this, even if practitioners have an intuition that "this is definitely stolen," they often get blocked because they cannot clearly present the grounds for their report.
The images owned by a brand are not just simple photos, but the brand's unique images—essential assets that represent intellectual property and build consumer trust. If these unique images are unauthorizedly copied, modified, and distributed, it does not stop at simple copyright infringement but shakes the brand's quality control and trust value. Therefore, from the brand's perspective, finding "a tool that can objectively prove even cleverly modified theft" is of desperate importance.
7 Representative Patterns of Clever Modifications
As mentioned earlier, the reason unauthorized sellers modify official images "just enough to not get caught" instead of copying them exactly is to ensure that objective evidence is weakened even if a report is filed. In particular, the problem is that they repeat modifications at a level that is difficult to detect by algorithms or simple comparison systems, even though consumers feel it is the same product.
Research findings also clearly show this vulnerability. According to a study presented at the 2022 USENIX Security Symposium, even simple modifications like cropping the image slightly, rotating the angle a bit, or adjusting only the brightness frequently caused existing image detection systems to mistake it for a different image from the original. In fact, these simple modifications alone resulted in a detection avoidance rate of over 99%. Another study (FAccT 2022) showed that even highly sophisticated systems like Apple's NeuralHash suffer from collisions where they fail to recognize the image as identical to the original simply by slightly distorting or rotating it.
Furthermore, counterfeit sellers were able to evade detection nets because adding minute noise to an image or converting it to grayscale and recoloring it changed the image fingerprint generated internally by the system to be completely different from the original. In fact, a 2021 study from the University of Illinois confirmed numerous cases where search technology using image fingerprints was neutralized by simple modifications like brightness adjustment, rotation, and cropping. It also showed that even global search engines like Google, Bing, and TinEye failed to link the modified images to the original, proving that search tools commonly relied upon by brand practitioners are not perfect solutions either.
The clever image modifications frequently identified in actual practice can be summarized into seven major patterns:
Changing Background Color/Tone: This involves changing a white background to gray, or adjusting the overall hue to be warmer or cooler. Fine-tuning the white balance is also used to make it look different from the original. Consumers recognize it as the same product photo, but differences from the original are created, making detection difficult with simple hash comparisons or surface-level similarity detection. Color and brightness changes are known to be sensitive points for traditional hashing or some scanning systems.
Cropping/Adjusting Margins: This method involves cropping the image slightly or increasing the margins to change the ratio. The product itself is exposed in the same way, but the overall structure of the image changes, making direct comparison with the original difficult.
Slight Rotation/Mirroring: This is when an image is rotated by just a few degrees or flipped horizontally. Consumers don't notice much of a difference, but existing simple matching algorithms may recognize it as a completely different image.
Moving/Concealing Logo Location: This involves covering or repositioning the brand logo to obscure the identifying points of the brand's unique image. The legal judgment of where infringement begins varies case-by-case and is not certain.
Covering Watermarks/Banners: This method involves removing or covering watermarks or promotional banners inserted by the brand as protective measures with other elements. Although it looks like simple editing on the surface, it constitutes an act of damaging the original protection mechanism.
Compositing (Bundles/Adding Props): This is the method of compositing other product images onto the same product photo, or adding props and text. In particular, if they switch the packaging to look like a 'bundled product,' proving the theft of the original becomes even harder.
Manipulating Resolution/Compression Rate: This involves saving the original image at a lower resolution or reducing the file size. Although there is no major difference to the naked eye, the accuracy of pixel-level comparison drops, which can lower the detection rate.
These modifications trigger unnecessary disputes and repetitive reviews during the reporting process, greatly consuming the resources of practitioners. Because the grounds are ambiguous, internal judgments are split, and unnecessary resources are wasted in communicating with platforms. Ultimately, it reveals that the core of the problem lies in 'how to turn intuition into persuasive evidence.'
How is AI Image Similarity Diagnosis Different?
We have already established that the 'partial modifications' made by unauthorized sellers weaken the persuasiveness of the evidence. The remaining task is to turn the intuition held by practitioners into evidence in a form that platforms can understand. Subjective judgments of simply being "similar" are hard to be accepted by platforms, and techniques like manual comparison or simple hashing are easily rendered useless in the face of clever modifications. How, then, can brands overcome this problem?
AI-based image similarity diagnosis bridges this gap. The key is turning human intuition into data. AI analyzes images at the pixel and pattern levels, presenting "where they are the same" and "how similar they are" simultaneously.
Heatmaps visually highlight the areas used as the basis for similarity. For example, they highlight with color the parts that humans might easily miss, such as label text lines, bottle shapes, or package patterns. This result goes beyond the simple intuition of being "similar" and can be utilized as visual evidence in internal reports or materials submitted to platforms.
Similarity Scores quantify similarity as a numerical value between 0 and 1. A higher score means a greater match with the original, but rather than concluding based on the number alone, its persuasiveness increases when interpreted alongside the Heatmap.
This analysis provides relatively robust results even against modifications (rotation, cropping, brightness adjustment, etc.) that simple comparison tools struggle to handle. Of course, since what kind of evidence each platform will accept is not disclosed and cannot be clearly defined, the ability to prove "where and how much they are the same" through visualization and numerical values is a great strength for practitioners. How to utilize Heatmaps and Similarity Scores in the actual reporting process, and what strategic interpretations are needed for each score, will be covered in more detail in the next content.
Ultimately, AI image similarity diagnosis does not just filter out simple duplicates, but provides a system that can prove even cleverly modified images with objective data. From a brand's perspective, this data enables faster internal consensus and decision-making, and allows them to present much more persuasive materials in external negotiations with platforms.
Conclusion: The Path to Protecting Brand Value, and the Preparation Practitioners Can Start Today
The conflict between brands and unauthorized sellers ultimately boils down to the question of "whether to protect the brand's unique image, or to cleverly exploit it." Brand companies invest massive costs to produce images, conveying consistent quality and trust through them. On the other hand, some sellers repeatedly engage in modified theft to reduce photography costs and gain quick exposure. The misconception that "it's okay if I change it a little" continuously fuels these attempts, but it eventually leads to consumer confusion and damage to brand value.
Therefore, simply preventing direct copies is not enough. An objective evidence system is needed to prove that even if they change the background color or move the logo, the original is still the brand's unique image. AI image similarity diagnosis complements this, catching clever modifications and turning the intuitive judgments of practitioners into a language provable with numerical and visual data.
To respond to this problem more effectively, it is important for practitioners to establish an evidence management system.
Official Image DB Management: Record the source, version, and posting date of original images
Capturing Suspected Images: Include the sales page URL, capture time, and seller ID
Storing Heatmap & Similarity Score Results: Prepare snapshots of AI analysis results as evidence
Creating Comparison Tables: Organize the original vs. suspected copy side-by-side to reduce internal alignment costs
Checking the Latest Platform Guidelines: Check report forms and the latest notices from platforms like Coupang and Naver
A brand's path to protecting its value from unauthorized theft starts with a response based on 'data and evidence,' not 'intuition.' There is no longer a need to remain in the frustrating situation of 'knowing it is stolen but finding it hard to prove.' We hope this article helps practitioners equip themselves with more objective evidence and prepare their first steps toward protecting brand assets.