Memorable Password Authentication

A conceptual photorealistic hero image blending memory and security: a glowing golden padlock rendered as a neural, brain-like network of light, floating above an open human palm in a dark, softly lit study. Faint bokeh of binary digits drifts in the background. Warm key light on the hand, cool blue rim light on the padlock, shallow depth of field, cinematic, high detail — evokes both 'memorable' and 'authentication' for a presentation title slide.

Michael Clark
Department of Computer Science
Brigham Young University

What Does it Mean to Authenticate?

A professional museum conservator authenticating an ancient large ceramic amphora in a laboratory. The expert is wearing white cotton gloves and examining the pottery carefully under a soft inspection lamp. The amphora is tall, weathered, with visible age cracks and subtle painted patterns. The setting is a clean conservation studio with wooden tables, archival boxes, magnifying tools, and documentation papers. The conservator is focused and serious, using a small magnifier to inspect surface details. Soft natural light from a window mixes with warm task lighting. Photorealistic, high detail, shallow depth of field, professional documentary photography style, 85mm lens, realistic skin texture, museum-quality scene.
Authenticate: prove something is real, true, or genuine.

Can prove membership of a class (broad, low specificity) or identity (high specificity).

Computer Authentication Factors

There are three basic ways to identify a terminal user.

  1. By something he knows or memorizes. He could memorize a password or answer a prearranged set of questions. This technique requires no special hardware and is reasonably secure.
  2. By something he carries. This might be a badge, card or key. The badge would be inserted into the terminal badge reader, the key into the terminal itself.
  3. By a personal physical characteristic. This might be the user's voice, which when transmitted to the computer would be compared with the stored "voice-print" for identification. This technique is under development, but not yet commercially available.

The Considerations of Data Security in a Computer Environment, IBM 1970

Authentication Factors Compared

Know

Advantages: always with you, easy to change, use requires personal consent, cheap/easy to implement

Disadvantages: memory, often weak, scales poorly, undetected theft

Have

Advantages: no recall, theft detection, unguessable, often phishing resilient

Disadvantages: misplaced/loss/theft, hard to share, cost per user, possible cost per account

Are

Advantages: always with you, can't misplace, no recall, stable

Disadvantages: privacy, not secret (NIST SP 800-63), involuntary change (ouch), no voluntary change, hardware reader cost

Authentication Tradeoffs

⚖️
  • "Better" depends on context
  • Why study passwords?
    1. Common: high leverage rewards
    2. Relevant: research attention
    3. Open: most research focuses on replacement
    4. Unique advantage: requires volition (until stolen)

Human Chosen vs Random Passwords

Human Chosen

Strengths:

  • Memorable
  • Comfortable

Weaknesses:

  • Zipf's Law
  • Targeted Guessing

Randomly Generated

Strengths:

  • No guess advantage
  • Known security (tunable)
  • Safer shorter

Weaknesses:

  • Less memorable
  • Algorithm aversion

Guessability, Language

A macro photograph of a handful of tumbling casino dice mid-roll against a dark background, motion blur suggesting randomness, with a faint overlay of glowing binary 0s and 1s in the negative space. High contrast lighting, shallow depth of field, photorealistic — conveys the concept of entropy and unpredictability.
  • Shannon Entropy = information, predictability, "bits"
  • Random values - IID (Independent and Identically Distributed)
  • Encryption keys, random passwords
  • Guessability = Entropy iff IID

Entropy Stretching

Hands stretch taffy against a tree branch bolted to the wall in a traditional manner, in what appears to be a low-tech industrial kitchen.
Stretching taffy. Cropped from https://wikitravel.org/en/File:Melcocha.jpg

Password → PBKDF (Password-Based Key Derivation Function) → Encryption key

  • "Key stretching" ≈ Proof of Work

Passwords let you authenticate, Cryptwords decrypt.

Authentication vs Encryption

Digital Authentication: Two or more parties. The subject (user) presents a proof of a credential to the relying party (server). With passwords, the credential typically is the proof. The relying party compares it to a verifier to authenticate or deny the subject.

Encryption: One or more parties. Using a secret or private key, reversible mathematical operations decrypt data from seeming-noise to the original data encrypted.

Two parties for authentication vs one for encryption or verifier guesses.

Password Reuse, Password Managers

A view of the KeePassXC password manager central database window.
KeePassXC Database View. From https://keepassxc.org/screenshots/

Server compromise interception + password as proof = no safe password reuse! (aPAKE + JS auditing)

Per site = password manager. Password manager = no memory burden = random generation.

Relying parties still MUST store secure verifiers instead of raw passwords:

  • credential stuffing
  • delay account compromise

Focus on entry usability instead of memory.

The End of Passwords?

A photorealistic flat-lay photograph of a laptop keyboard, a smartphone showing a PIN entry screen, and a physical numeric PIN pad, arranged together on a plain desk surface. Soft, even studio lighting from above, neutral gray background, clean commercial product-photography style.

Password manager = no memorized passwords?

Authenticate: computer, phone, workstation, debit card

Password manager primary password

Real choice!

Memory matters.

How can we make randomly generated passwords maximally memorable?

What Helps Memory?

A photo of a man with a furrowed brow, straining to remember. Bokeh background of a testing center. Intricate detail.

Memory aids:

  • Primacy / Recency
  • Length
  • Cues
  • Recognition
  • Pictures > Words
  • Incongruence / Schema violation
  • Own choices (even guided)
  • Chunking
  • Repetition (spaced)
  • Depth of processing
  • Focused Time
  • Mnemonics
  • Sleep
  • Strong Emotion

Intent to recall does not help for random sequences.

How Long: Authentication Passwords

✋

Two parties; either can cancel.

  • Rate limits / Lockouts
    • PINs: 4 digits, since 1966
  • "Unlimited" attempts; Denial of Service
    • 46.6 random bits = 3/10,000 odds after 1,000 guesses/second for 1 year
    • 47.6 bits: 8 character mixed case alphanumeric

How Long: Cryptwords

A young girl shrugging, with the caption, 'Why don't we have both?'
https://knowyourmeme.com/memes/why-not-both-why-dont-we-have-both

277 bits

embargo flint canon requisite pointed bargraph unicycle likely unsaved jasmine selective chloride unfair unsorted graduate pedicure buddhist squishier nail skirmish gangrene deflector

dqEDJkJjX­YFoXYRQRR­9dDjvXHnr­UqRjmiqXg­ipK2yh97­4RU

Let's trade less usability for security.

Minimal sufficient length.

Cryptword Relevance

Computers guess faster every year. Human recall more stable.

Cryptword lengths a moving target; become obsolete?

Cryptwords: Moving Window Work Factor

Graph plotting encryption key strength, extra work, and cryptword length. There is a constant size gap between encryption key strength and extra work. There is a flat line labeled Cryptword Strength, at 64 bits in this figure. Thin dotted lines connect encryption key strength and extra work at 2000 and 2030.
Using extra work, cryptwords can have constant length despite computational advances

PBKDFs: tunable work factor

Increase work as hardware improves.

Requires:

  • Active developers
  • Software updates
  • Enter cryptword to update protection
  • Keep verifier secret, rotate everything if secrecy lost
    • No storage-free password managers
  • Attacker capability grows in pace with consumer capability

Same length = same strength, forever.

Michael Clark and Kent Seamons. Passwords and Cryptwords: The Final Limits on Lengths. New Security Paradigms Workshop (NSPW), 2022.

But Can People Remember Long Passwords?

A screenshot the 3D SeedQuest interface. A dinosaur park environment. In the lower right of the screenshot is a quarter circle containing an image of a pterosaur labeled 'Pterosaur', and below that, 'screeching'.

SeedQuest. 128-bit passwords, mapped to actions in a virtual space.

Positives:

  • Method of Loci inspired
  • Spatial position (context cue)
  • Chunking/batching
  • Bizarre Imagery (incongruent)

Negatives:

  • Selection with replacement
  • Three tiers of choices (scene, object, state)
  • Repeated elements between scenes - yeti

Test Function and Cause

A screenshot the 2D ordered SeedQuest interface. Pictures of interactable elements are positioned in place over a picture of a dinosaur park environment. In the lower right of the screenshot is a quarter circle containing an image of a pterosaur labeled 'Pterosaur', and below that, 'screeching'. A screenshot the 2D random SeedQuest interface. Pictures of interactable items are presented in two rows of 8 each over a dinosaur park environment. In the lower right of the screenshot is a quarter circle containing an image of a pterosaur labeled 'Pterosaur', and below that, 'screeching'.

Test if it works and how:

  • If: Passphrase vs SeedQuest - within subjects
  • How: between subjects, to avoid fatigue.
    • Hypothesis: 3D is same as 2D
    • Hypothesis: Position memory helps
  • Single shot, 7- and 30-day follow-up.
  • Mechanical Turk (N=300 → 208 → 144)
IRB Approval seal

SeedQuest Results

Graph showing two pairs of lines. The lines labeled 'passphrase' drop to 0 in about half the time as the lines labeled 'path' A graph plotting participants as dots in three columns, the height of the dot (0 to 100%) showing how high a percentage of the path they recalled. There are diamonds overlaid on the dot columns; a bar in the middle represents the group mean (~58% for 3D, ~45% for Ordered, ~40% for Random), while the top and bottom represent the upper and lower confidence interval of the group mean respectively. A black horizontal line shows the overall mean (~45%). The height of the diamonds overlap, indicating no statistical significance. A histogram to the right side shows somewhat even spread, with a large bump at 0% for Ordered and Random.

10% perfect recall, but more of path recalled.

No significant advantage of 3D (p=0.0878); post-hoc analysis revealed 3D helped recall scenes specifically.

Michael Clark, Scott Ruoti, Michael Mendoza, and Kent Seamons. A Comparison of Three Approaches to Assist Users in Memorizing System-Assigned Passwords. Symposium on Usable Security and Privacy (USEC), 2024.

Password Choice

A screenshot of the password generator interface from GeneratePasswords.org. The left side of the screen is a list of controls, and the right side has passwords. The left starts with two tabs labeled 'Password' and 'Passphrase'; 'Password' is selected. Below that is a slider labeled 'Length' with range from 4 to 32, and to the right of the slider is a numeric input with the value '8'. Below the sider is a section header 'Include', followed by 5 checkboxes labeled 'Include numbers [0123456789]', 'Include lowercases [abcde…xyz]', 'Include uppercase [ABCDE…XYZ]', 'Include symbols [?-*\%!@#_\$:;/]', and 'Include brackets (\[\]\{\}()<>)'; the first three checkboxes are selected and the last two labels are struck through. Below these is a heading labeld 'Exclude' followed by 2 checkboxes labeled 'Exclude similar characters i, I, l, L, 1, o, O, 0' and 'Exclude unwanted characters'. Neither is selected, and both have a help icon to the right. On the right side of the screen, skipping past irrelevant heading information (the site name, the page header 'Generate secure passwords', etc), there is a strength meter with 3 of 5 bars filled in and the label 'Reasonable'. Below that in small font it says '47.63 bits' with a help icon to the right. Below that is an eye icon labeled 'hide' on the left, and on the right a highlighted button labeled 'Refresh'. Below that is a list of random passwords; 7 are show in the screenshot, though a scrollbar to the right indicates there are dozens more. Each password has a 'copy' icon to the right.

Lists of random passwords are common. Choice helps recall.

  • Do the lists help?

Most studies are of complete systems, not individual elements!

Design of Experiment

Screenshot of a website. It shows a list of 9 passwords, with a cursor hovering over one highlighted in blue. Above, the text 'Thank you for participating in our study. We have generated 128 passwords for you to choose from. Please select a password of your choice from the provided list.'
  • 4 groups, between-subjects: 1, 8, 32, 128 passwords to pick from
    • 200/group to detect 10% improvement, 25% dropout expected
    • Prolific: 1066 → 859 → 626
    • 7- and 28- day follow-up
IRB Approval seal

Security

Screenshot of a website. It shows a list of 9 passwords, with a cursor hovering over one highlighted in blue. Above, the text 'Thank you for participating in our study. We have generated 128 passwords for you to choose from. Please select a password of your choice from the provided list.'
  • Vary dictionary: lowercase, base57 (no "l1I0O"), full alphanumeric
  • Equal security: longer for longer lists; potential perfect predictions
Password character counts
Choices
Dictionary1832128
Lowercase7888
base576677
Alphanumeric6667

List Lengths Results

A bar chart of Choices on the x axis and Percent on the y axis. Choices has 4 bins: 1, 8, 32, and 128. Each of the 4 bins has two bars, labeled 'FALSE' and 'TRUE'; the legend is labeled '7-day Recall'. Percent ranges from 0 to 60. The true bars are all around 60 and the false bars all around 40, though the last bin slightly differs, at around 52 true and 48 false. A bar chart of Dictionary on the x axis and Percent on the y axis. Dictionary has 3 bins: lowercase, Reduced Alnum, and Full Alnum. Each of the three bins has two bars, labeled 'FALSE' and 'TRUE'; the legend is labeled '7-day Recall'. Percent ranges from 0 to 60. The true bars are around 55 to 62 on the Percent axis, while the false bars are around 45 to 38 on on the Percent axis. The middle true bar, 'Reduced Alphanumeric' is slightly higher, but they are all roughly the same height.
While keeping strength constant, no evidence that memorability increases as much as 10% when changing list lengths.

Dictionary didn't matter, as shown in "Correct Horse Battery Staple".

Michael Clark, Gregory L. Snow, and Kent Seamons. Choose from a List: A User Study of Random Password Memorability. CHI Conference on Human Factors in Computing Systems, 2025.

Can We Learn From Password System Studies?

Problem:

  • Password system studies combine many factors

But:

  • Many such papers
  • Semi-consistent reporting

Statistics to the rescue!

Labeling Studies (Qualitative Coding) - Deductive

Theory-driven codes - discussed memorability research before collection

  • Recall % ⭐
  • Follow-up delay
  • N (participants)
  • Test type (free/cued/recognition/skill)
  • Who chose secret (user/random/assigned)
  • Graphical (picture superiority)
  • Mnemonic device
  • Spaced repetition

Finding Papers (Literature Search)

Labeling Studies (Qualitative Coding) - Inductive

Additional Codes: Weekly meetings, new codes discussed/added

Exclusions: no memory outcome, or no reported "success rate"

LASSO

Linear regression: estimate impact on outcome while controlling for other variables!

LASSO (Least Absolute Shrinkage and Selection Operator): variable selection and regularization

What Matters Most?

A mess of lines branching off from a horizontal line at 0 on the y axis. The y axis is labeled coefficient, and the x axis Penalty, with ticks decreasing left to right logarithmically from 5000 down to 2. The figure looks vaguely like an inside-out umbrella or tree roots if rotated 90 degrees. On the right side are labels for each colored line; they read from high to low: Test: Mixed, Test < 1 day, Test: Recognition at around y=1, Mnemonic Devices, Memory Check, Crowdsourced, Choice: User, Graphical, Practice Reps, Test:Cued Recall, N at y=0, Choice: Random, Test 1-3 days, log(N), Choice: Assigned, Choice: Mixed, Test: Free Recall, Test > 1 week, Test 4-7 days, Chunked, Test: Skill, and Mnemonic Phrases at around y=-1. The position of the labels indicates their estimated normalized effect size in a model with all labels included; where they enter indicates they are relevant enough to overcome that much of a penalty for complexity. The first to enter is Practice Reps, then Choice: User, then N (which is replaced later by log(N)). The first two strong ones through the rest to enter are Test < 1 day and Chunked, one positive and one negative.

Additional Findings

Michael Clark, Trevor Bond, Gregory L. Snow, and Kent Seamons. Password Memorability: What Matters Most? SOUPS Poster Session, 2025.

Overall Takeaways

Passwords can remain relevant.

Well known memory aids help best:

Chunking performed much worse than expected - future work should validate this.

Relevant Future Work

How do key memory design choices impact each other?

Authentication Factors reclassification

How long in the wallet to memorize passwords and cryptwords?

Can consumers create cryptwords?