Cybersecurity
Behavioral biometrics - how to deploy continuous session verification
Behavioral biometrics in banking and in the enterprise, step by step: purpose, GDPR and impact assessment, AI Act, its role next to MFA, alert responses and a mimicry test.
Józef SulwińskiPublished Updated 9 min read

Contents
A bank customer logs in with a password and a second factor, and then someone else uses the session. A one-time check at login will not notice. Behavioral biometrics compares the way a person types, moves the mouse and handles the phone with the account owner’s pattern for the whole session.
NIST guidance, an EBA opinion and EU law show that:
- NIST calls the continuous evaluation of a session session monitoring and lists typing cadence among the signals a system may evaluate,
- in its 2019 opinion, the EBA treats keystroke dynamics and the angle at which a device is held as possible elements of strong customer authentication for payments,
- the GDPR counts behavioural characteristics as biometric data when they allow or confirm the unique identification of a person.
What behavioral biometrics measures
The system learns how a specific person uses a device and compares current behavior with that pattern. The result of the comparison is a probability of a match with the pattern.
| Signal | What the system compares |
|---|---|
| Keystroke dynamics | The system compares typing speed and the intervals between key presses. |
| Mouse movement | The system compares the cursor path, its acceleration and the way the user clicks. |
| Touchscreen use | The system compares screen pressure and the way the user swipes. |
| Holding the phone | The system compares the angle at which the user holds the device. |
NIST lists these characteristics as examples of behavioral biometrics. It also points out that biometric characteristics are not secret and that their comparison is probabilistic, unlike checking a password.
Two ways to use it
| Way | What it does | What to watch |
|---|---|---|
| Authentication factor | The system checks the pattern at login or when a payment is authorised. | The EBA treats such a factor as compliant only when the implementation gives a very low probability of authenticating an unauthorised party. |
| Session monitoring | The system evaluates behavior for the whole session after login. | NIST notes that most signals have privacy implications, so their collection must go through a privacy risk assessment. |
A bank that wants to count behavioral biometrics as an element of strong customer authentication must demonstrate the quality of its implementation. According to the EBA, the quality of the implementation decides compliance.
Deployment step by step
Step 1. Write down which event the system must detect and what should happen then.
An example event is a session taken over after a correct login, or a transfer ordered by someone other than the account owner. The security team records the event, the signals and the expected response in one document. Why: without a written purpose, thresholds cannot be set and nobody can check after deployment whether the system works.
Step 2. Establish with the data protection officer whether the system processes biometric data within the meaning of the GDPR.
The GDPR defines biometric data as data from the technical processing of physical, physiological or behavioural characteristics that allow or confirm unique identification. Biometric data used to uniquely identify a person belongs to the special categories of data. Its processing is prohibited unless one of the exceptions in the regulation applies.
The controller records the legal basis and the exception it relies on. Why: without this decision, the deployment may be unlawful on the day it starts.
Step 3. Assess the impact on data protection before the system goes live.
Under the GDPR, the controller assesses the impact of processing that uses new technologies when it is likely to result in a high risk to people. The assessment is always required for large-scale processing of special categories of data. Why: the assessment shows which data is really needed, how long to keep it and who may access it.
Step 4. Check which AI Act category the system falls into.
The AI Act classifies remote biometric identification as high-risk. It excludes verification whose sole purpose is to confirm that a person is who they claim to be. A recital adds that biometric systems used solely for cybersecurity and personal data protection are not high-risk.
A system that also assesses the user’s emotions, for example stress, falls into a different category. Emotion recognition is high-risk, and inferring employees’ emotions in the workplace is prohibited, except for medical or safety reasons. Ask the vendor for a written description of what the system derives from the signals. Why: the same technique can be ordinary identity verification or a prohibited practice.
Step 5. Use behavioral biometrics as an additional layer next to multi-factor authentication.
Multi-factor authentication (MFA) requires at least two kinds of proof from the user, for example a password and a hardware key. Do not replace MFA with behavioral biometrics alone. Why: NIST allows biometrics only as part of MFA with a physical authenticator, because biometric comparison alone does not give enough confidence.
Step 6. Write down the response to each alert level.
NIST lists three responses to suspected fraud during a session: reauthentication, terminating the session and notifying support personnel. A rule can look like this: if a signal appears before a transfer is ordered, the system requires reauthentication. If the signal repeats, the system ends the session and notifies the fraud team.
Why: a team without a written response decides on it during the incident, while the customer waits.
Step 7. Prepare a path for people rejected in error.
Biometric comparison is probabilistic, so the system will sometimes treat the account owner as a stranger. A rejected user gets another way to confirm their identity, and the team records every such case. Why: the number of false rejections shows whether the thresholds are right, and the customer keeps access to the account.
Step 8. Test resistance to mimicry.
In 2018, researchers at the University of Waterloo described an attack on keystroke dynamics verification on a smartphone, in which an augmented reality (AR) app guided the attacker in mimicking the victim’s pattern. With its help, participants carried out more than 400 attacks, and 87% of them bypassed verification after an average of four minutes of training. During the deployment test, try to mimic a user’s pattern after observing them. Why: only a test on your own application shows how the system copes with mimicry.
Step 9. Do not use voice to verify identity.
Why: NIST SP 800-63B-4 prohibits biometric comparison based on voice. A phone call with a customer needs another method of confirming identity, for example a confirmation in the bank’s app.
How to assess a vendor’s solution
| Question | What to check |
|---|---|
| Which signals does the system collect? | Obtain the full list of signals and compare it with the purpose recorded in step 1. |
| What does the system derive from the signals? | Check in writing that the system does not assess emotions or sensitive characteristics. |
| Where are the patterns stored? | Establish the storage location, the retention period and who has access to the patterns. |
| How does the system respond to an alert? | Check whether responses can be set separately for login, payments and changes of personal data. |
| How often does the system reject account owners? | Ask for test results and compare them with your own pilot. |
| How does the system cope with mimicry? | Check whether the vendor has described tests of mimicry attacks, and repeat them on your own application. |
Checklist
- The event the system must detect and the expected response are recorded in one document.
- The data protection officer has established whether the system processes biometric data and has recorded the legal basis.
- A data protection impact assessment was completed before the system went live.
- The system’s AI Act category is recorded, and the vendor has confirmed that the system does not assess emotions.
- Behavioral biometrics works as an additional layer next to MFA.
- Each alert level has a recorded response: reauthentication, session termination or notifying the team.
- People rejected in error have another way to confirm their identity, and their number is measured.
- The deployment has passed a pattern mimicry test on your own application.
- Voice is not used to verify identity.
Where we help
During penetration tests of web and mobile applications, we check session management and authorisation logic. We establish whether a hijacked session allows an operation to be ordered and whether the responses described in steps 6 and 7 work. In bank implementation projects, a security architect reviews the authentication design and the choice of vendor. You are welcome to reach us through the contact form.
Legal basis
| Step / topic | Where it is |
|---|---|
| Behavioral characteristics as biometrics, limits of biometrics, ban on voice comparison | NIST SP 800-63B-4, 3.2.3 (ban on voice comparison: 3.2.3.2) |
| Session monitoring, signals and responses | NIST SP 800-63B-4, 5.3 |
| Behavioral biometrics as an element of strong customer authentication | EBA-Op-2019-06, paras 18-19 and Table 1; RTS under PSD2, Art. 8 |
| 2. Definition of biometric data | GDPR, Art. 4(14) |
| 2. Special categories of data and exceptions | GDPR, Art. 9(1) and 9(2) |
| 3. Data protection impact assessment | GDPR, Art. 35(1) and 35(3)(b) |
| 4. Biometric identification and verification | AI Act, Art. 3(35) and 3(36); Annex III point 1(a); Recital 54 |
| 4. Emotion recognition | AI Act, Art. 5(1)(f); Annex III point 1(c) |
| 8. Mimicry attack on keystroke dynamics | Khan, Hengartner, Vogel, MobiSys 2018 |
Related articles
- Password security - stolen credentials and authentication methods
- Zero Trust Architecture - continuous access verification in IT and OT environments
- Data privacy - GDPR in practice - impact assessments and legal bases for processing
Sources
- NIST SP 800-63B-4: Authentication and Authenticator Management - NIST (accessed: )
- EBA-Op-2019-06 - Opinion of the European Banking Authority on the elements of strong customer authentication under PSD2 (21.06.2019) - European Banking Authority (accessed: )
- Commission Delegated Regulation (EU) 2018/389 (regulatory technical standards on strong customer authentication) - European Commission (accessed: )
- Regulation (EU) 2016/679 (GDPR) - EUR-Lex (accessed: )
- Regulation (EU) 2024/1689 (AI Act) - EUR-Lex (accessed: )
- Khan, Hengartner, Vogel - Augmented Reality-based Mimicry Attacks on Behaviour-Based Smartphone Authentication (MobiSys 2018) - ACM (accessed: )
Need help in this area?
Our experts will help you assess the risk and plan next steps.