An experimental method for CSS based fingerprinting and a pure CSS 'supercookie'.
- Contents
- What is it?
- How does it work?
- CSS Cookie
- Why is this important?
- Examples
- Calculating Device Uniqueness
- Improvements and Further Research
- Contributing
- License
CSS Fingerprinting is a technique of tracking and gathering information on site visitors. This method exploits the nature of CSS to collect various characteristics about the visitor's browser and device, which can later be used to either identify or track said visitor.
By sending a variety of media queries that apply to specific browser characteristics, the browser will select a set of styles that apply to itself. We then trick the browser into sending this information back to the server by setting the background-image of these styles to a specific URL. The server will then respond with HTTP Status 410 (Gone) to avoid any requests of these characteristics on subsequent reloads.
For example, to detect the type of pointer input:
.pointer {
background-image: url('/some/url/pointer=none');
}
// Coarse (touchscreen)
@media (any-pointer: coarse) {
.pointer {
background-image: url('/some/url/pointer=coarse');
}
}
// Fine (mouse)
@media (any-pointer: fine) {
.pointer {
background-image: url('/some/url/pointer=fine');
}
}Installed fonts can also be detected in a similar manner:
@font-face {
font-family: 'some-font';
src: local(some font), url('some/url/some-font');
}
.some-font {
font-family:'some-font';
view raw;
}However, this works a little differently; every font not installed on device will send a request. By comparing the differences between the requests and the full list of fonts, we can conclude what fonts are installed.
We can also track visitors cross-origin by requesting an endpoint on the server that will return a permanent redirect (HTTP status 308) to a unique address. The browser will then permanently make requests to the previously generated unique address whenever the endpoint is requested. This creates a pure CSS cookie that is reminisent of the 'supercookie' exploit. This cookie is stored for an unlimited amount of time; the only way to remove it is to fully clear the browser's cache.
This technique avoids anti-tracking methods such as NoScript, VPNs or browser extensions, as it requires no Javascript or Cookies to function.
Currently, this method is not scalable as it requires over 1MB of CSS downloads and hundreds of requests per user. However, with the next upcoming draft of the CSS specification, CSS Values 4, it may dramatically shrink the number of requests per user by allowing the use of custom variables in URLs.
.body {
--unique-identifier: 'foo'; // unique generated ID
--pointer: 'none';
--theme-preference: 'none';
// Only make one request
background-image: url("/some/url/?" + var(--unique-identifier) + "&" + var(--pointer) + "&" + var(--theme-preference));
}
// Detect pointer type and theme
@media (any-pointer: coarse){
body {
--pointer: 'coarse';
}
}
@media (prefers-color-scheme: dark) {
body {
--theme-preference: 'dark';
}
}Not only will the upcoming draft make this method scalable, but it will also increase its precision. Currently, without alternative means, it is hard to conclusively link every request to a specific visitor as the only feasible way to determine their origin is to group the requests by the IP address of the connection. However, with the new draft, by generating a randomized string and interpolating it into the URL tag for every visitor, we can accurately identify all requests from said visitor.
Included in this repository you will find an implementation of CSS Fingerprinting using the old method, fingerprint.sass, and an example of how to instantiate it, example.sass.
You can find examples of different css-tracking servers in the examples directory.
To see a complete example (HTML/CSS/Server) check out the study repository.
Shannon Entropy is used to quantify how identifiable fingerprint is. Let H be the entropy, X a discrete
random variable with possible values {x1,..., xn } and P(X) a probability mass function.
Shannon Entropy takes the following formula:
The entropy of Shannon is in bits where b = 2. One bit of entropy reduces by half the probability of an event occurring.
Rust
type Fingerprint<'a> = Vec<(&'a str, &'a str)>;
type DataSet<'a> = Vec<Fingerprint<'a>>;
fn shannon_entropy(data: DataSet, value: usize) -> f64 {
let key_occurances = data
.iter()
.flatten()
.fold(HashMap::new(), |mut acc, &(key, _)| {
*acc.entry(key).or_insert(0) += 1;
acc
});
let kv_occurances = data
.iter()
.flatten()
.fold(HashMap::new(), |mut acc, &(key, val)| {
*acc.entry((key, val)).or_insert(0) += 1;
acc
});
let mut entropy = 0.0;
for kv in &data[value] {
let p = *kv_occurances.get(kv).unwrap() as f64 / *key_occurances.get(kv.0).unwrap() as f64;
entropy -= p * p.log2();
}
entropy
}A set of performance and accuracy improvements that could be made to the method.
Whilst many privacy browsers such as Brave will attempt to mask the use of NoScript to avoid fingerprinting,
these attempts could be thwarted by applying styles that will only be rendered in noscript tags:
<noscript>
<p style='background-image: url("/some/url/noscript=true")'>
NoScript Detected
</p>
</noscript>
Currently the fingerprinting.sass example will test all values between one and an arbitrary limit.
This method is highly inefficient and a little inaccurate.
For example: CSS pixels in actuality are split into fractions when resolved by the browser and hence two devices with similar, but non-identical dimensions will be counted as the same.
Not only does this method cause inaccuracy but it also is inefficient. Most devices can be grouped into categories of similar dimensions. In the case of phones and tablets the differences between their dimensions will be extremely small and a higher accuracy is needed to identify the differences. However, there large gaps in size between the different groups (for example between Tablet and Desktop resolutions), which means there is little need for accurate testing between those ranges.
Further Research: Determine the optimal precision parameters for both intra and inter group testing.
Through the use of Javascript, we can do several things to improve the accuracy and performance of this technique:
- Delayed/Async Loading - By delaying the loading of fingerprinting files with JS, we can ensure that the browser loads the rest of the page before these files, improving page responsiveness.
- Sharding - By splitting the fingerprinting files into component groups we can again reduce the performance cost by downloading them in parallel.
- Conditional Execution - Sharding the files, also opens the possibility of conditional execution. If a fingerprint can be uniquely identified by a subset of shards, there is no need to burden the server with the overhead of loading the full set.
Further Research: Develop standardised sharding and conditional execution practices to improve performance and reduce server load.
Most operating systems ship with a certain set of default fonts and display configurations. By testing a subset of known default fonts that are included on different operating systems, we can, with a certain degree of confidence, determine which OS is installed on the device.
If this could be implemented it would dramatically reduce the number of requests per user as font-detection is the most costly part of the process.
Further Research: Determine a standardized subset to test for. This set should test for the key differences between the defualts of different operating systems.
If user generated CSS is displayed on websites it may give attackers the ability to track other visitors.
If you have any problems, changes or additions, please just open an issue or pull request!
All content is licensed under the MIT license and is purely for educational purposes.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent