BMI Biometrics: What do we know?

Just like fingerprints, brainprints do have inheritable characteristics. However, unlike many other inputs to biometric analyses (fingerprints, irises, face-photos), brainprints have tremendous immunity to spoofing.

A brain-machine interface (BMI) is an instrument that records changes in electrical potential associated with thought, motor movement, or blood flow within a person’s skull [1,2,3]. As every person’s brain develops differently, these electrical potentials can be analyzed to distinctly identify persons with high accuracy. This is particularly relevant to the field of biometrics, in which biological indicators, such as fingerprints, faces, DNA, or brainwaves, can be used to discriminate between individuals. Today, biometrics are used for authentication when passwords may not be optimal, such as when someone may be “shoulder-surfing”, or if insider-threat level is high [4].  In BMI biometrics, the feature most often (but not always) studied is the “brainprint” — an underlying set of distinct features emitted from a person’s skull during a mental task. Given the remarkable applications of a one-person-one-vote layer to the internet, brain-based biometrics’ distinct advantages have high-potential applications.

It may come as a surprise that the earliest relevant work in BMI biometrics had the intention of finding genetic similarities, not identifiable differences. As early as 1936, researchers demonstrated that brainwave recordings were more similar among identical twins than among two random persons [5]. Three decades later, it was concluded that crude brainwave features have an autosomal-dominant method of inheritance [6]. It wasn’t until 1999 that the first person-identification study using brainwaves was published.  In (Poulos, 1999a), alpha-band rhythms of brainwave recordings identified individuals in a group with 72-80% accuracy [7]. It took only a few months for Poulos to apply a new method to the same dataset, improving his classification results rate to 95% through convex polygon intersections [8].

In 2007, the first large-scale BMI-biometric study was conducted by Palniappian et. al. By analyzing subjects’ brainwaves as they memorized images from a black-and-white picture set, a classification rate of 98.12% in 102 subjects was obtained [9]. In 2010, Brigham et. al applied support vector mechanics on brainwave recordings during subjects’ imagined speech, resulting in a 98.96% classification accuracy in 120 subjects, alerting the security community that brainprints were here to stay [10].

Just like fingerprints, brainprints do have inheritable characteristics. However,unlike many other inputs to biometric analyses (fingerprints, irises, faces), brainprints have tremendous immunity to spoofing. An attacker cannot synthetically generate BMI recordings to assume the identity of another individual as they can with a fingerprint [11], iris, gait [12], or facial recognition system. This is due to the tremendous complexity of the underlying biochemical processes of the brain and lack of public data available for training spoofers. When obtained in a human-detectable setting, brainprints may also be immune to stealing even from spoofing algorithms, unlike DNA. This may be performed in the short-term using vybuds as part of a “time-dependent input, biology-dependent output” communal hash, and in the long-term using OpenMined‘s mantra of “if good data were easy to fake, machine learning theorists would fake data instead of collecting it”.

In addition to BMI’s built-in “liveness” detector, BMI biometric systems are unique in that they require a user’s willingness to cooperate. Because brain patterns are heavily influenced by mood and stress-level [13,14], identification is not valid when forced by an outside party. This is a major benefit for user-safety. While this volition requirement is true for the vast majority of constructed paradigms for BMI biometrics to date, in some special cases (such as proving that a particular person committed a crime), a volition-insensitive BMI biometric may be desired. Initial research into developing protocols where identification is consistent regardless of a person’s stress-level has been explored[15].

Another advantage unique to BMI biometrics is their mobility. Recordings can be made unobtrusively, allowing for continuous verification throughout a secure task. This prevents one-time login attacks and person substitution. With continuous verification, a log-in cannot be transferred to others to allow them access to sensitive data. Given that digital security systems are most often thwarted by insiders [5], this advantage is particularly enticing for those working with corporate enterprise. Likewise, new research is taking this distinct feature of BMIs to the next level. Continuous verification allows for users to subtly initiate emergency broadcasts, without alerting a threatening individual who may be watching (see later section –  “Security Applications: Beyond Biometrics”). These “covert warnings” make use of distal artifacts (such as clenching one’s teeth) to secretly call for help without alerting one’s attackers that it has been requested [14], and are only possible through a continuous verification protocol.

A final dose of idiosyncrasy offered by BMIs is their inclusiveness. Those with severe injuries such as burned or missing fingers, aniridia (absence of the iris), or severe paralysis (such late-stage ALS or “locked-in” syndrome) are not excluded from using a BMI biometric. Due to their inclusiveness, BMIs are considered ‘universal’.

A number of different modalities have been used to record brainwaves, but to date, only two have been used in person-identification. The first, and most relevant, is the electroencephalography (EEG). The second, discussed later, is functional magnetic resonance imaging (fMRI).



  5.       D&d – Need to find this citation still again
  6.       – Lost this citation too
  13.   S. Marcel, J. R. Millan. Person authentication using brainwaves (EEG) and maximum a posteriori model adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(4):743–748, April 2007.


  4.   – Need to find citation for this one as well  
  5.   Auditory potentials –!etd.send_file?accession=ucin1439300974&disposition=inline
  9.   Facial movements, eeg authentication using artifacts
  10.   Multi-level approach based on eye-blinking
  12.   DOI is good
  13.   DOI vs fMRI –


Leave a Reply

Your email address will not be published. Required fields are marked *