What a keyboard sees
On Android a keyboard is an input method: a separate app that every other app hands its text fields to. Messages, searches, addresses, card numbers and passwords all pass through it before the app they were meant for receives them. That is why Android shows a warning when you switch on a keyboard you installed yourself, saying it may be able to collect everything you type.
The keyboard that came with your phone gets no such warning. It was chosen for you, it is often a system app you cannot uninstall, and its settings are the defaults almost nobody changes. That is what makes collection by it passive: nothing has to be agreed to, enabled, or noticed.
Who marks a field private
A keyboard is told what kind of field it is typing into by the app. A password field declares a password input type. Separately, an app can set a flag asking the keyboard not to learn from what is typed there, which browsers use for private tabs.
Both are requests, not enforcement. A keyboard that wants the words can ignore them, and one that honours only the no-learning flag still learns from password fields, because Android’s own text fields do not set that flag for passwords.[6] Whether your passwords end up in a prediction dictionary is decided by keyboard code you cannot read.
Telemetry, by default
The most careful public study of mainstream Android keyboards we know of is a 2022 Trinity College Dublin dissertation that captured and decoded the network traffic of Google’s Gboard, Microsoft’s SwiftKey and the open-source AnySoftKeyboard.[1] Gboard is the keyboard preinstalled on Pixel phones and many others.
Both proprietary keyboards sent large amounts of telemetry: timestamped entries with hardware details, the length of each word typed, the languages used, and the name of the app the keyboard was open in. Gboard’s logs carried the phone’s Android ID and SwiftKey’s carried the Google Advertising ID, so the records tie to one handset. Gboard offered an opt-out; SwiftKey did not. The study notes that timestamped app usage alone can support inferences about the person typing.
It is just as important to say what the study did not find. Neither keyboard was seen sending the text itself, or even per-character counts. AnySoftKeyboard sent no telemetry at all, which the author reads as proof that this collection is a choice, not a requirement.
Cloud prediction
Some keyboards do send what you type, by design: they ask a server to predict the next word. Then the question becomes how well that traffic is protected.
In 2023 Citizen Lab found that Sogou Input Method, with over 450 million monthly users, sent keystrokes under a home-made encryption scheme that a network eavesdropper could break, on Windows and Android.[3] In 2024 it tested the cloud-prediction pinyin keyboards of nine vendors, including the ones phone makers ship: Samsung, Xiaomi, OPPO, vivo, Honor and Huawei, plus Baidu, iFlytek and Tencent. Eight of the nine had flaws that could expose keystrokes in transit; only Huawei’s held up. Combined with Sogou, the researchers put the exposure at up to one billion users.[2] Most vendors fixed the issues after disclosure, and the reports concern eavesdroppers on the network, not what the vendors themselves do with the text.
Updates and servers
A keyboard that talks to a server needs that channel to be safe in both directions. In 2015 NowSecure disclosed CVE-2015-2865: the SwiftKey-based keyboard Samsung preinstalled fetched language-pack updates over plain HTTP without checking them, ran with system privileges, could not be disabled, and kept checking for updates even when another keyboard was in use. NowSecure’s conservative estimate was 600 million affected devices. SwiftKey’s own store apps were not affected; the flaw was in how the keyboard was built into the phone.[4]
And data collected by a keyboard company lives on that company’s servers. In 2017 Kromtech found an unsecured database belonging to ai.type, a popular third-party keyboard rather than a manufacturer’s, holding records on about 31 million Android users, including data drawn from contact lists. The company disputed how sensitive it was.[5] The general point stands without the dispute: data that is never collected cannot be left exposed.
What we do instead
FluxBoard removes the places this data could go rather than promising to treat it well.[6]
- No analytics, crash reporter, advertising library, account or Google Play Services. CI fails the build if a Play Services, ML Kit, Firebase or
datatransportclass appears in a release APK, with a positive control so a broken scan fails loudly. - Prediction is on the device. Nothing about typing touches the network; the only network use is moving your library to another phone of yours over your own Wi-Fi.
- Password and PIN fields are detected from the field’s own input type, not the app’s flag. There the keyboard stops learning, suggesting, autocorrecting, decoding glides and capturing the clipboard. A manual privacy switch covers the fields Android gives no signal about.
- The personal dictionary and clipboard history are excluded from Android’s cloud backup, and on Android 12 and later, cloud backup is refused outright on a phone with no lock screen.
What we haven’t proven
The checks above read the source, the dependency tree and the built APK. Nobody has yet captured FluxBoard’s live network traffic on a physical phone, the way the studies above did for other keyboards, and reading code is not a substitute for watching it run.[6] The privacy policy says so too. If you do that capture, we would like to see it.
FluxBoard is also a beta keyboard with English only. Everything above is about where your data goes, not a claim that it types better.
Sources
- K. Przepiorowski, supervised by D. Leith, "How Private Are Android Keyboards?", MCS dissertation, Trinity College Dublin, 18 April 2022.
- Citizen Lab, "The Not-So-Silent Type: Vulnerabilities Across Keyboard Apps Reveal Keystrokes to Network Eavesdroppers", 23 April 2024.
- Citizen Lab, "Please do not make it public: Vulnerabilities in Sogou Keyboard encryption expose keypresses to network eavesdropping", 9 August 2023.
- SecurityWeek, on NowSecure’s disclosure of CVE-2015-2865 in the SwiftKey-based keyboard preinstalled on Samsung phones, June 2015.
- The Hacker News, on Kromtech Security Center’s finding of an exposed ai.type database, December 2017.
- FluxBoard, docs/PRIVACY.md, for every claim about FluxBoard here.