PimEyes Face Search Engine Tested
PimEyes offers an online face search engine that has attracted significant attention, but how well does it work?
IPVM tested the features and capabilities of PimEyes through a series of tests. In this report, we show our results:
- How does PimEyes' performance compare to Google's?
- Are social media images included in PimEyes results?
- What percentage of search results are explicit?
- Does PimEyes show duplicate or unique images?
- How do PimEyes matches compare to open-source face recognition?
- What is the timeframe on alerts?
- How do PROtect Service and GDPR Takedowns work?
- What are the advantages of "Deep Search" to standard search?
- How well does PimEyes detect faces?
- What is the minimum pixel density?
- Plus 2 videos showing how PimEyes works and the results of matches against security industry executives.
Test ***********
** **** *******, ** *********** * dataset ** *** ********* ***** ****** four ********* ****** (**, ****** *****, Security, *** ******).
*** ** ******* ******** ***** *********** by*** ********* ************ *** ** ***** ****** ** male *** ****** ***** ****** * variety ** ***** *** ******.
*** *** ****** ***** *******, ** selected ***** **** ********* ****** ****** media ***** (****** ** *** ****,****** ** *****, ********* ** ******).
******** ******** ********** *********** *** ******** dataset:
*******, ** ******* * *** ** politicians *** ******* *** *** ****** dataset:
Executive *******
** *** *******, ******* ********* *** superior ** ****** (********* ******* ***** search *** ********* ** ******'* ****) returning *** **** ******* **** ******. The ****** ** ******* ** **** commonly ******** ** *** ******** **** of *** ****** ***** ******** *** in *** *****, ******* ******** ************* more ******* *******.
*******, ******* ******* ****** **** ********* results (*** **** ****** ** * person) *** ******** ***** (***** *** some *****, *.*., ******* ****, ** a *******). ****** ****** **** * 7-10 *** *****, *** ************ ******** matches ** *** ***** ****.
*** *** ** ******* ** *******, however, ** ** ******** ** ******* that **** **** ****** *****, ** line **** *** *******'* ******** **** it **** *** ***** ***** ***** (e.g., **, *******, *********, ******). ******* these ******* (*.*., ***** ********* ****), the ******* ****, ** *** **** of *** ********* ****** ****** ** the ********).
** ***** *** *****-******* ******* ***** we ****** * **** ***** *** not *******, ********** **** ******* ********* may *** ***** *** ******* ********.
** ****** *** ***-*** ******* *** it ******* **** *** *** *** of *** ****** ** ****** *******, which ** ******* ** ******* ******* did *** ***** ***** ****** **** a **** ****** ********** ** *** opt-out ******* *****.
Video ** *** ******* *****
Security ****** ******* *********
****** ******* ******** **** ******* *** the ******** ********** *** ***** *****.
**** ** *** ******* ******* ***** were **** ****, ****** ******* **** publishes ** ********* ** ******** **********.
PimEyes ******** ****** **** ****** ******* ***** ******
** ********* ******* ***** ******** *** our ******* ***** ******* *** ******. For ***** ********* ** **, ******* identified ~*** ******* ** *******, ******* Google ******* ***** ****** ******** **** than * ***** *** ****. * minimum ** * ******* *** * maximum ** **** ******* **** ********* by ******* *** *** ** **-********* faces.
*** *** ****** ********* **** ****** media, ******* ******* ~*** ****** ** average, ******* ****** ********** * *****. PimEyes ******* ****** **** * ** 2716 ******* *** ****.
******** *** *** *** ******** ********** returned ~*** ******* ** ******* *** a ****** **** * ** ****** on *******. ******* ******** ** ** 338 ******* *** *** ******** *******.
*******, ** ******** *** ****** ** famous ******. ** *******, ******* ******** ~1,420 ***** *******, ******* ****** **** showed ~**. *** **** ***** ** faces, ******* ******* ****** **** *** to ****. *** ******* ****** ** results ******** *** ****** ****** *** much ***** ** ********** ** *** maximum ****** ** ******* **** *** AI-generated ***** (**** ******** ** ****). One ***** ****** ****** ****** ** have ******* **** ** ********* ** images ******, *** ******* ******* ********** around ~*,*** ****** *** "*-****" *********** and ****-***** ***********.
No ****** ***** ****** ******** ** ******* *******
********* *** ****** ******* *** *** dataset, ** ****** ***** ***** **** detected. **** ******** *******' ****** ** not ******** ****** ***** ***** (********, Instagram, ********, *******).
**** ************ *** ******.*** ***** ** the ************** ****** ***** ******** *** did *** **** ****** ******* ************ targeted ****** *******. *******,************************** **** "*** *** ** ****** or ***** ********* ***** ** ****** without ******* ********** ** ******** **********."
More ******** ***** ******** *** ****** ********
*** **-********* ******, ***** *** * significant ********** ** ******** **** ********** across ********* *******, ***** ******* ******** more ******** ****** *** *******. **** was **** **** *** *** ****** Media *****. ** ******** ******* **** found *** *** ******** *****, *** less **** *% ** ***** **** identified ** ******** *** ****** ******.
*** ********** ****** ******* ** ** images *** ****** ***** ******** ***** raise **** ********, *******, ** ** not ******** ** ****** ** ******* targets ******* *** ******** ******** ** if ***** *** *** **** ******** images ** ******* ** *** ***.
*******' ********* ********** ************* *** **** ******** ** ********* are *** "******* ****". **** ******** the ***** ** ****** ******* ** identify ******** *****. *** ********* ******** for ******* ******** (***, ******, ****, xxx, *****, ******, *****) ****** *** weblinks ** ********* ** **** **** was *** ******** ******* ** ***. At *** **** ** *** *******, PimEyes *** *** **** * ****** result ** * "*********** ******** ******", but **** *** ******** * ******, added **** *************.
Duplicate ****** *** ***** **** ********* ********
****** ****** *** *** ****** ********** of ****** ******, **** **** **** 6% ** ******* ** *******. ** our ***********, ******* *********** *** ****** directly ******** *** ***** ****, **** searches *** ******.
** ************* ********** ********* ****** ***** the***-**-******** ******. ******, ** ********* ******** ** the ****** ***** *******-** ****** ******** (SURF) *** **** * ******* ********** of *** ********* ********. ** **** clustered *** ******** **** * ****** vocabulary, *********** ********* ********* ****** **** the ***** *****, *** ******* ****, until **** ****** ****** ********.
PimEyes ******* **** ***** **** ****-****** **** ***********
** ******* ******** *** ******** ******* of *******, ** *** * ****** recognition ********* **** *** ****** ******. Particularly, ** *******-****** *********** ********* ****** ** ******* ****** *********** on *** ***** ******** ********.
*** ****** ** ***** ******* ******* VGG-Face *** **** **** **** ******* outputted. ** **** *****, **** *** not **** ****** ** ******-******* ***** but **** ********* ***** ***** ***** indexed ** *******.
Alerts ** ***** **** **** * *****
******* ****** ***** ** *** * reference *****, *** *** ****** ** new ******** ****** ***** ***** **** subscription. ***** ******* *** ****** ** the ********* *** ******* ***** ******.
** **** ** *** *******, ** published******** ****** ** ****, ***** ******* ****** ******* **, ****. ******* ******* the ****** ** *** **** ******* on ******* **, ****, ***** ******* to ******** *** ***** ** *** other ****. *** *** ***** ****** on *** ******** **********, ** *********** a ******* ********* ** *-** **** for ******* ** ***** *** ******.
"PROtect" ******* *** **** ********* *** *** **** *********
***** *** ***** ** **** ** remove *** ******** ****** ****** *** PimEyes ******** ***** *** "*******" *******. In ***** ** ****** *** ******, ID ************ ** ********, *** *** only ** **** ** *** ****'* personal ******. ***********, ***** *** ****** a ******** ******* ******* ******* ** the ******* ******.
**** *** **** ******* * ******** request, ******* ****** * ******** ****** on ****** ** *** **** *** sends ** ** *** ******* ** question. **** **** *** ********* *** removal ** *** ***** **** *** website, *** ** ** *** ***** example, ******* ********* **** *********** ************* ** ****** *** *****.
"Deep ******" ******* ***** *** ******* ** ******** ******
******* ****** ********* ****** **********, **** as ******** *** **** ***** *** the ******, **** ****** ****** (******** warnings *** ******** *****), *** **** search ******. **** ****** ** **** available *** *** **** ********* ************ tier (********), *** ** ** *** clear ** ******* ************* **** ** the ************ ********* ** **** ******.
** ****** ** ****** ******* ** compare *** ****** ** ****** ******* between ******* *** **** ********. ** average ******* ****** ******* ** ~**% of *** ******* ********* ** **** search. *******, **** ******* ********** ******* depending ** *** ****** ** **** search *******. *** **** ******** **** 1,000 ** **** *******, ******* ****** only ******* ~%** ** *** *******, whereas, *** **** ******** **** **** than *,*** *******, *** ******* ****** returns ~%** ** *** *******.
Good **** ********* **** **********, ****, *** ********* ******
******* *** **** ** ******** ***** at ********* ****** (****** **/**** ** sideways, ***-**** ****** *****), **** ********** and ****, *** ** ****** **** multiple *****.
Missed ******* *** ************* ********
** *** ***** ****, ******* ******* images **** *** *** ******* *** target *******. **** ** *** ******* could ** ********** *************, *** **** missed ******* *** *** **** * close ***********.
Minimum ** *** ********
**** ****** ******* ***** * ******* of *** *** ****-******* ******, *** 85 *** *** ********** ** ** the ****** ******* ** ************ ****** a **** ** *** ***** *** perform * ******. ******** ***** ** an *****, *** *********** (*******, *****, etc.) **** ****** **** ********* ** perform ******.
*** ******* **** ******* ******* ********* their *** ********:
**, ******** ** *** *** ********, there *** *******: *. ****** ****** now **** **** ******* ***** *********** to ****** *********** ** ***** ******** results. *. ** *** *** **** interface *** ****** ******* *** *** be ******* ** ********** **** ********** to *** ******* ****** *****. *. There *** **** ******* *** ******* in ***** **** *********, ** **** it **** **** ********. ** **** fixed *******/********* ********* *** ***********. *. As *** ******** ******, ** **** change *** ***-*** ******. *** ******* it ***** **** ********** ***** **%. 5. **** ** **** ******* ** test *** ********** ******, ***** **** automatically ****** ********** ******** *** ***** potential **** (** *** ** ***** with ******** **** ** ** ****).
******* ****** "** ** *** **** accurate *** *******" ****** ****** **** recent *** ******** *******.
******* ********* * *** ******** ******* *, ****, "********* ****** mechanics" *** "********* *** ******** **************." The *** **** *** ****** ******** improvement ** "*** ********** ******, ** reduce ***** ******** *******" *** "*** results *** ******* ********* ** **********."
***** ** **** ******* ** *** update **** * ***** ****** ****, the ******* ****** *********, *** **** doppelgangers *** ***** *****.
**** **** ****** **** ****** **** any ******* ***** *******.