3.9 KiB
3.9 KiB
Abusive Account Detection
Helpful bunnylol
| aad | aad wiki pages |
| go aadata | abusive accounts data |
| fblearner | model trainings are listed here. |
| orb | inmemory DB |
Data
| ig_signup_sigma_features | |
| ig_challenge_.... | accounts that were challenged |
proxy metrics
| enrolments | tested users (UFAC) |
| clear | the lower better (UFAC cleared) |
human labeling
| holdout | signed users, after 12days we label them by humans -> empty, bad, good (bennain) ~5K avccounts |
false negarive
MAU prevalense MIMA prevalence
Folders
-
Misc
-
fbcode/dataswarm-pipeline/tasks/si/fake_accounts -
fbcode/dper3/dper3_models/si/olf -
www/flib/intern/scripts/sigma/clssifiers/olf -
configerator/source/sigma/online_classifiers/runtimes- all our classifiers are here
-
configerator/source/si/fake_accounts- defines active classifiers and the defaults
-
si_sigma/Lib/FakeAccounts- sigma rules for the fake accounts, namely new_user_registration is processed here
-
-
Models
fbcode/fblearner/flow/projects/fluent2/domains/si/aad_surfaces- fblearner models
-
Sentry
configerator/source/si/sentry/prod/<namespace>/<category>.cconf- configuration of sentries.
- e.g.
namespcae=facebook, category=new_user_registrationdefines what is passed to sigma rules. This is used for FB. IG has different config perhaps.
configerator/source/si/sentry/si_namespaces.thrift- sentry namespaces
www/flib/si/sentry/category/SentryCategory.php- existing categories
www/flib/si/sentry/preparable/filters/sigma/SigmaFilter.php- Sigma filter that can be found in sentry configuarations.
- bunnylol orb
- this can be used to query sentry logs
-
Sigma
si_sigma/Endpoint/Sentry/SentryFollowProfile.hssi_sigma/Contexts/Sentry/SentryFollowProfile.hs- scuba sigma_profiling
- to profile
-
QE
configerator/qe2_diff/newExperiments/vahagnk_fast_tiger_clone.txt
-
Reg Attack
configerator/source/si/reg_attacks/surface_definitions.cinc- This here we define surfaces.
configerator/source/si/reg_attacks/attack_definitions/- Attack definitions.
source/si/reg_attacks/reg_attacks.thrift- FieldTypes are here.
-
Piplines
-
OLF
- phps OLFAdminV2 status --classifier reg_enthusiastic_impala
- firefighting
Model Training workflows
- Train the model
flow-cli canary si.olf.ig_signup.train@olf --run-as-secure-group=team_abusive_accounts_detection --entitlement si --parameters-file configs/ig_signup_andromeda_offline.json
to monitor progress use bunnylol fblearner.
- Publish model
aimps publish-model --manifold --oncall aad_surfaces --is-dper-model -d service_sharded <model_id>_<snapshot_id>
to monitor progress use bunnylol predictor.
- Register model
phps --www-root /var/www OLFAdminV2 training --action=register_model --classifier-name <model_name> --problem IG_FA_ANDROMEDA --surface IG_SIGNUP --model-id <model_id> --threshold 0.5
model name is ig_signup_colorful_animal
- Compare how model fires.
- Create experiment to compare enrolled vs cleared. bunnylol qe2
Team identifiers
- fawg - this is abusive account detection group for diffs