Just thought I’d post a heads up on a few techniques actively being used to circumvent copyright violation detection. I’m tired of finding stolen designs and faked images every time I load a page on here. I figure posting this in the open will be a wash - more folks will know how to do it and more will know what to look for and report.
The newer techniques are:
- Using video frame grabs and thumbnails (cropped and whole).
- Color replacement and object cloning/reordering - Take a product image with a widget in 3 colors, copy-paste a few, adjust the hue of the pasted objects, and optionally adjust the hue of the whole image.
- Composition, especially with adding text. Instead of just using the stolen image, they expand the canvas, add a second image with a matching background or a call-out bubble, and add some text.
Those are the newer ones, designed to look different from the original (in case they’re near each other on search results) and avoid reverse image search detection. The old ones were simple, single operations - crop, adjust brightness or saturation, add a banner, mirror the image, etc.
The old ones could be detected easily with a reverse image search. I use a Firefox plugin that throws a list of reverse image search engines in the right click menu (called Reverse Image Search).
The new ones are tougher, but I’ve found techniques that help.
First, try the easy way. Check for other people’s logos, text that is/isn’t there, and search here or other model sites for a keyword and skim the results. Reverse image search using the highest resolution variant you can find of the full image. From there, try cropping out clutter and added text. From there, try selecting only a unique aspect of the object.
If that doesn’t work, you’ve got to try tricking the search engines to examine features, but with minimal information on the actual object. The theory is that a low level of confidence in the object’s identification will filter out a large number of shopping links. If the engine can identify the object with high confidence, it will overfit the general object to present you with lots of click-through shopping bait. You want the object identification to be fuzzy so that the engine doesn’t get creative. Try:
-
Limiting the selected area in the results to a single object instance.
Example: Image has 5 different colored vases - select only 1 vase. -
Limiting the selected area in the results to unique aspects of the photo or everything but the actual object.
Example: If the item is depicted on a desk, limit the analyzed area to any visible clutter. If a hand is holding it, search for the hand and a smidgen of the object (enough to get some of the object’s color). -
Limiting the selected area in the results to a narrow band.
Example: Object is a decorative mask - select only the vertical band center between the mask’s eye holes or only from the outside edge of an eye hole to the edge.
Finally, I’ve found Google and Bing to complement each other well. Different data sets, different techniques, so you get different results (especially when your selection is very uncertain).
Once you’ve found one suspicious post, check the account’s full portfolio. It’s rare that I find a single violation. It’ll usually be that the account has a few super-basic designs and a bunch of rip offs. Just report the whole account.
I tend to peruse the listings sorted by newest and find 5-ish fakes per 50 posts. That usually leads to 5-ish additional fakes per account.
Anyhow, yeah, it’s disappointing when you start to actually investigate. I tend to do my little “audit” every few months here and at Printables. I’ll see an obvious rip off and get curious, then wake up from my trance 30 minutes later to 20 tabs open and a partially completed report or 3.
