SLIDE 1
David Bounie, Prof. david.bounie@telecom-paris.fr
1
Télécom Paris, Institut Polytechnique de Paris ww.xai4aml.org
The team
2
David Bounie, Economics Winston Maxwell, Law Stéphan Clémençon, Applied Mathematics Astrid Bertrand, PhD
David Bounie, Prof. david.bounie@telecom-paris.fr Tlcom Paris, - - PowerPoint PPT Presentation
David Bounie, Prof. david.bounie@telecom-paris.fr Tlcom Paris, - - PowerPoint PPT Presentation
David Bounie, Prof. david.bounie@telecom-paris.fr Tlcom Paris, Institut Polytechnique de Paris ww.xai4aml.org 1 The team David Bounie, Economics Winston Maxwell, Law Stphan Clmenon, Applied Mathematics Astrid Bertrand, PhD 2
SLIDE 2
SLIDE 3
The partners
SLIDE 4
Alert
FINANCIAL INTELLIGENCE UNIT FIRST LEVEL REVIEW SECOND LEVEL REVIEW AUTOMATED REVIEW SYSTEM: deterministic or improved by machine learning
The AML process in a nutshell
4
LAW ENFORCEMENT Customer & transaction data, sanctions data Complementary sources*: Web, social media, newsfeed…
Alert confirmed Suspicious Transaction Report Referral to juridical authority
DATA GATHERING
Bank processes FIU Law enforcement
SLIDE 5
99% of alerts have little or no utility to law enforcement
5
TMS generates 100 alerts
Out of 100 alerts, 10 are converted into SARs (>90% false positives) FIUs analyze only 2 SAR cases out of 10 FIUs forward 1 to law enforcement Bank processes Law enforcement processes
SLIDE 6
€200 billion in criminal funds circulate every year in Europe… …Yet only 1% of criminal funds are confiscated.1
1 Europol, “From suspicion to action - Converting financial intelligence into greater operational impact.”, Sept. 2017.
6
SLIDE 7
Current AML systems exhibit high false positive alerts while showing low success rates in catching illicit funds.
7
AI can help improve AML systems, but creates new forms of risks for individuals. Explainability is a key factor to mitigate those risks.
SLIDE 8
Alert FINANCIAL INTELLIGENCE UNIT FIRST LEVEL REVIEW SECOND LEVEL REVIEW AUTOMATED REVIEW SYSTEM: deterministic or improved by machine learning
How AI can help?
8
LAW ENFORCEMENT Customer & transaction data, sanctions data Complementary sources*: Web, social media, newsfeed… Alert confirmed Suspicious Transaction Report Referral to juridical authority Customer segmentation Anomaly detection Prioritization of alerts* Vizualisation tools* * Not Always deployed by banks. AML stages where Machine Learning can be valuable DATA GATHERING
SLIDE 9
Why is explainability important?
9 SOCIETAL POINT OF VIEW
AML-CFT: Why is it suspicious enough to shut down? Fundamental rights: Why are my assets being frozen?
TECHNOLOGIC AL POINT OF VIEW
Is my algorithm functioning properly? Do I detect any drift in behavior over time?
REGULATORY POINT OF VIEW
Is my model AML compliant?
Contrarily to other profit-maximizing firms that will always prefer efficiency over explainability, Financial Institutions won’t risk a fine for AML, a field that has no business value.
SLIDE 10
Multiplicity of audiences:
Internal controller, Auditor, ACPR, TRACFIN, Individual impacted by the model, CNIL.
Challenges for explainability in AML
10
CHALLENGES FOR EXPLAINABILITY IN AML
Multiplicity of business cases: bank/insurance Avoid bias & discrimination Access to the data :
Data used for the explanation can raise data privacy issues.
Economic viability :
The explanation should also take into account economic factors.
R&D
Go beyond post-hoc approaches (LIME, SHAP, etc.): Graph Conv. Networks, etc.
SLIDE 11
Thank you for your attention!
Point of contact: david.bounie@telecom-paris.fr Research Chair website: www.xai4aml.org
11