Corso di digitalizzazione e reti per segretario amministrativo
Intelligenza Artificiale e Sicurezza Informatica
1. Intelligenza Artificiale nella PA: esperienze e
prospettive
Roma, 5 Novembre 2019 h. 9.00-13.30
organizzato e promosso da
www.prsmonitor.euwww.di.uniroma1.itwww.miur.gov.it
Intelligenza Artificiale e Sicurezza
Informatica
Prof. Luigi V. Mancini
Dipartimento di Informatica
Sapienza Università di Roma
11. Sapienza is among the Members of the Algorand
Foundation university committee;
12. organizzato e promosso da
12
Nominativo Speaker
Ruolo - Organizzazione
Dijsktra Prize 2019 – Prof A. Panconesi
13. Success of Deep Learning
• Branch of machine learning that makes use of neural networks, to find solutions for a variety
of complex tasks either in supervised or unsupervised way
- Areas used:
- Computer vision
- Image processing
- Face recognition
- Speech recognition
- Text-to-speech systems
- Natural language processing
- Games...
13
image source: https://goo.gl/u26HM3
image source: https://goo.gl/xNwTVw
image source: https://deepmind.com/research/alphago/
14. organizzato e promosso da
14
Nominativo Speaker
Ruolo - Organizzazione
Attacks
Machine Learning to perform Attacks
Deep Fakes and Social Bots
16. organizzato e promosso da
16
Nominativo Speaker
Ruolo - Organizzazione
Social engineering
Companies are prey to increasingly sophisticated
computer scams, carried out using social engineering,
emails, etc.
With deep fake, you could get to simulate a
videoconference from the CEO of the company.
17. organizzato e promosso da
17
Nominativo Speaker
Ruolo - Organizzazione
The Building block of DeepFakes
The Autoencoder
18. organizzato e promosso da
18
Nominativo Speaker
Ruolo - Organizzazione
How DeepFakes work? First phase
19. organizzato e promosso da
19
Nominativo Speaker
Ruolo - Organizzazione
Cont... 2nd phase
20. organizzato e promosso da
20
Nominativo Speaker
Ruolo - Organizzazione
Test di Turing
A machine considered intelligent if it is indistinguishable from a human being.
29. Attacks on ML models
1) Hacking Smart Machines with Smarter Ones, 2011 by Mancini et al.
2) Model Inversion Attacks, 2015 by Fredrikson et al.
3) Membership Inference Attacks, 2017 by Shokri et al.
29
32. organizzato e promosso da
32
Nominativo Speaker
Ruolo - Organizzazione
Concluding Remarks
Researchers are constantly engaged in exploring Machine Learning and
the related Privacy and Security Issues.
In order to find innovative and effective solutions, it is necessary to invest
in research and not only in the market.
Acquire Kow-how so we can always be a little ahead of the attackers.
The Public Administration protects the interests of the community and
cannot afford to lag behind the attackers.
34. G
D
34
Discriminator D
Eyewitness
Generator G
Police Sketch Artist
Dataset
“Real”
Data Sample
“Fake”
Data Sample
real
sample
fake
sample
GAN Architecture
Generative Adversarial Network (GAN) basic
architecture