Tag Archives: HTML

The ability of websites to bypass privacy settings with “canvas fingerprinting” has caused quite a bit of concern, and it’s become a hot topic on the Code4lib mailing list. Let’s take a quick look at it from a technical standpoint. It is genuinely disturbing, but it’s not the unstoppable form of scrutiny some people are hyping it as.

The best article to learn about it from is “Pixel Perfect: Fingerprinting Canvas in HTML5,” by Keaton Mowery and Hovav Shacham at UCSD. It describes the basic technique and some implementation details.

Canvas fingerprinting is based on the <canvas> HTML element. It’s been around for a decade but was standardized for HTML5. In itself, <canvas> does nothing but define a blank drawing area with a specified width and height. It isn’t even like the <div> element, which you can put interesting stuff inside; if all you use is unscripted HTML, all you get is some blank space. To draw anything on it, you have to use JavaScript. There are two APIs available for this: the 2D DOM Canvas API and the 3D WebGL API. The DOM API is part of the HTML5 specification; WebGL relies on hardware acceleration and is less widely supported.

Either API lets you draw objects, not just pixels, to a browser. These include geometric shapes, color gradients, and text. The details of drawing are left to the client, so they will be drawn slightly differently depending on the browser, operating system, and hardware. This wouldn’t be too exciting, except that the API can read the pixels back. The getImageData method of the 2D context returns an ImageData object, which is a pixel map. This can be serialized (e.g., as a PNG image) and sent back to the server from which the page originated. For a given set of drawing commands and hardware and software configuration, the pixels are consistent.

Drawing text is one way to use a canvas fingerprint. Modern browsers use a programmatic description of a font rather than a bitmap, so that characters will scale nicely. The fine details of how edges are smoothed and pixels interpolated will vary, perhaps not enough for any user to notice, but enough so that reading back the pixels will show a difference.

However, the technique isn’t as frightening as the worst hype suggests. First, it doesn’t uniquely identify a computer. Two machines that have the same model and come from the same shipment, if their preinstalled software hasn’t been modified, should have the same fingerprint. It has to be used together with other identifying markers to narrow down to one machine. There are several ways for software to stop it, including blocking JavaScript from offending domains and disabling part or all of the Canvas API. What gets people upset is that neither blocking cookies nor using a proxy will stop it.

Was including getImageData in the spec a mistake? This can be argued both ways. Its obvious use is to draw a complex canvas once and then rubber-stamp it if you want it to appear multiple times; this can be faster than repeatedly drawing from scratch. It’s unlikely, though, that the designers of the spec thought about its privacy implications.