Algorithm transparency in SourceAFIS

Algorithm transparency in SourceAFIS is an API that lets applications capture all intermediate data structures that are created during feature extraction and matching, including images in various stages of filtering, extracted minutiae and ridges, pairing tree constructed by matcher, and score breakdown. Learning basics of how the algorithm works is recommended before using transparency API.

Why?

The simple API exposed by SourceAFIS is generally an advantage, because the job of SourceAFIS is to abstract away details of fingerprint matching. There are however occasions when seeing what the algorithm is doing is beneficial. Specifically, algorithm transparency can be used to:

• explain how the algorithm reached its decision,
• create rich visualizations of fingerprints and their matches,
• check quality of captured fingerprints,
• use fingerprint features creatively for something other than matching,
• support fingerprint experts in manual matching,
• reconstruct damaged fingerprints in forensics,
• diagnose issues with sensors, outlier populations, and algorithm itself, and to
• export data for use in another biometric system.

Transparency data is also used internally to ensure consistency of language ports of SourceAFIS and to perform basic sanity checks in unit tests.

Algorithm transparency is rare in commercial matchers, becasue it weakens intellectual property protection. SourceAFIS, being fully opensource, doesn't need to hide anything from you. Algorithm transparency is its unique advantage.

Standard API data

Transparency data does not include data that is available in some other way. Specifically, it excludes the original fingerprint image, source string during deserialization, the resulting template (which is documented elsewhere), and the final score.

Full log

The simplest way to obtain transparency data is to ask SourceAFIS for a transparency ZIP file. The following example logs data from feature extraction, matcher initialization, and a single match.

FingerprintTemplate candidate = ...;
byte[] probeImage = ...;

try (
    OutputStream stream = new FileOutputStream("transparency.zip");
    FingerprintTransparency transparency = FingerprintTransparency.zip(stream)
) {
    FingerprintTemplate probe = new FingerprintTemplate(
        new FingerprintImage()
            .dpi(500)
            .create(probeImage));
    new FingerprintMatcher()
        .index(probe)
        .match(candidate);
}

See: FingerprintTransparency.zip() (Java)

This will create file transparency.zip populated by numerous JSON and binary data files. You can customize what gets logged by adding or removing method calls inside the try block.

Zip contents

List of files in the generated transparency.zip is shown below. Files named pairing and score appear multiple times, each containing different minutia pairing. Some of these repeats have been omitted from the list. Click on any file group to learn about its format.

Selective logging

ZIP file might be easy to generate and examine by hand, but most use cases for transparency API require programmatic access to data. While you could save the ZIP file and then load parts of it, it is easier and more performant to define your own FingerprintTransparency implementation like in this example.

// extend class FingerprintTransparency and define capture() method
class TransparencyContents extends FingerprintTransparency {
    @Override protected void capture(String keyword, Map<String, Supplier<byte[]>> data) {
        for (String suffix : data.keySet()) {
            byte[] content = data.get(suffix).get();
            out.printf("%,9d B  %-6s %s\n", content.length, suffix, keyword);
        }
    }
}

FingerprintTemplate candidate = ...;
byte[] probeImage = ...;

// use the newly defined logging class to capture data
try (TransparencyContents transparency = new TransparencyContents()) {
    FingerprintTemplate probe = new FingerprintTemplate(
        new FingerprintImage()
            .dpi(500)
            .create(probeImage));
    new FingerprintMatcher()
        .index(probe)
        .match(candidate);
}

See: FingerprintTransparency.capture(String,Map) (Java)

This should result in the following output. You can see the data is more structured. While this example just reports data length, it does load all serialized data into the content variable, ready to be parsed. Data is passed to the logger in the exact order in which file entries appear in the ZIP file above.

       24 B  .json  version
1,160,896 B  .dat   decoded-image
      161 B  .json  decoded-image
1,160,896 B  .dat   scaled-image
      161 B  .json  scaled-image
    1,518 B  .json  block-map
  665,600 B  .dat   histogram
          .   .     .
          .   .     .
          .   .     .
   13,514 B  .json  pairing
      734 B  .json  score
       87 B  .json  pairing
      642 B  .json  score
   17,579 B  .json  pairing
      725 B  .json  score
       17 B  .json  best-match

If you don't pull content from the Supplier, it will never get constructed. This lets you access algorithm transparency data with minimal overhead, picking only data you actually need.

Deserialization data

Above examples capture the primary workflow of feature extraction and matching. It is also possible to collect transparency data from template deserialization.

FingerprintTemplate template = ...;
byte[] serialized = template.toByteArray();

try (
    OutputStream stream = new FileOutputStream("transparency-deserialized.zip");
    FingerprintTransparency transparency = FingerprintTransparency.zip(stream)
) {
    new FingerprintTemplate(serialized);
}

The above code will produce archive transparency-deserialized.zip with the following contents.

This ZIP archive doesn't include the template itself, because serialized templates can be examined directly by following template format documentation.