4. 2. Experiment 2: retention over time

An authentication mechanism is only useful if authentication can still be accurately performed at some time after the password is memorized. In Experiment 2, we confirmed that sequence-specific knowledge acquired by users was retained over prolonged periods of time. Although skill learning generally persists over time, a SISL-based test had never been conducted with a substantial delay and a sufficient number of participants.

In Experiment 2, participants agreed to perform the SISL task over two sessions. In the first session, participants completed a training sequence with the same structure as the one in Experiment 1. The training was immediately followed by the same SISL test to assess sequence knowledge before the delay. A group of 32 participants returned to the online applet after 1 week to perform a retention test and recognition assessment for the trained sequence. A separate group of 80 participants returned after a 2-week delay for the retention and recognition tests. For the 1-week group, the test session consisted of a 540-trial implicit sequence learning assessment. For the 2-week group, the test session was doubled in length to additionally evaluate whether a longer test provided better sensitivity to individual sequence knowledge. For both groups, the initial speed of the test on the delay session was set to match the speed with which the participants had been performing the task at the end of the training session. A short warm-up block of 180 trials was used to adjust this initial speed so that participants were performing at around the target 70% correct at the beginning of the retention test.

Figure 4 shows gradual learning of the trained sequence
during the first session for both groups as in Experiment 1.
Implicit sequence knowledge at both immediate and
delayed tests is shown in Figure 5. On all five assessments,
participants exhibited reliable sequence learning as a
group, ts > 4. 3, ps < 0.01. On the 1-week delay test, 15 of 32
participants exhibited individually reliable sequence knowl-
edge. However, for the 2-week delay group, 49 of 80 partici-
pants exhibited reliable sequence knowledge, reflecting the
increased sensitivity in the longer assessment test used.
Future research will examine both increased training time
and 70.6% correct for the untrained sequence. The differ-
ence of 8.6% correct (SEa 2.4%) indicated reliably better per-
formance for the trained sequence by a paired-sample t-test
versus zero, t( 34) = 3. 55, p < 0.01.

Group-level differences in performance are commonly seen on tests of implicit learning, but being able to reliably assess individual learning is necessary for an authentication method. On an individual participant basis, performance on the trained sequence could be discriminated from that on the untrained sequence on the 540 test trials (by chi-squared analysis at p < 0.05) in 25 of 35 cases. For authentication purposes, the individual reliability of the assessment will need to be further improved by longer training to establish the implicitly learned sequence. However, the ability to identify learning in a large fraction of individuals with relatively short training is a feature of the SISL task not seen in most tests of implicit learning. 7 Traditionally, measures of implicit learning rely on assessing performance within groups of individuals, without the ability to identify learning at the individual level. 9

Explicit recognition test. After the training and test blocks, participants were presented with five different animated sequences and asked how familiar each looked on a scale of 0 to 10. Of the five sequences, one was the trained sequence and the other four were randomly selected foils. This test assessed explicit recognition memory for the trained sequence.

On the recognition test, participants rated the trained sequence as familiar at an average of 6. 5 (SE 0.4) on the 0– 10 scale and rated novel untrained sequences at 5. 15 (SE 0.3). The modestly higher recognition of the trained sequence was reliable across the group, t( 34) = 3. 69, p < 0.01, but did not correlate with SISL performance (r = 0.13), indicating that it did not contribute to the implicit performance. Slightly higher recognition of the trained sequence is often seen in implicit learning experiments as healthy participants find some parts of the training sequence familiar after practice. It is worth noting that implicit memory does not transform into explicit knowledge, even with repeated use, and the structure and length of the training and test sequences specifically aim to reduce the possibility that explicit knowledge is accumulated over time.

The general small difference in recognition ratings ( 5. 15 vs. 6. 5) indicates that participants would not be able to recall the 30-item sequence, meaning that they could not consciously produce the training information (e.g., to compromise the security of the authentication method). We discuss the reconstruction question further in our third experiment.

a SE is short hand for Standard Error. In other words, if the percent correct measurements for trained and untrained sequences followed the same normal distribution, the t-value calculated with N = 35 samples (and thus N − 1 = 34 degrees of freedom) should be near zero—less than the t- value obtained here of 3. 55, which represents a 99% probability that the difference is significant. The t-test is a standard statistical method used to confirm that the manipulated variable (here, sequence type) affects the measured variable (percent correct).

20

15

10

5

– 5

0 Tra ine d se qu enc ea d van ta ge

( %)

Training Block (540 trials)

1234567

1 week 2 weeks

Figure 4. Across training participants gradually begin to express knowledge of the repeating sequence by exhibiting a performance advantage for the trained sequence compared to randomly interspersed noise segments. Learning performance was similar across both groups and similar to Experiment 1, as expected.