• Federated Learning ๋‚ด ๋ง˜๋Œ€๋กœ ์ •๋ฆฌ

    2021. 11. 22.

    by. ํ•ด๋Š”์„ 

    Federated Learning : ์—ฐํ•ฉ ํ•™์Šต

    ๋‹ค์ˆ˜์˜ ๋กœ์ปฌ ํด๋ผ์ด์–ธํŠธ์™€ ํ•˜๋‚˜์˜ ์ค‘์•™ ์„œ๋ฒ„๊ฐ€ ํ˜‘๋ ฅํ•˜์—ฌ ๋ฐ์ดํ„ฐ๊ฐ€ ํƒˆ์ค‘์•™ํ™”๋œ ์ƒํ™ฉ์—์„œ ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋Š” ๊ธฐ์ˆ .

     

    ๋Œ€๋ถ€๋ถ„์˜ DL/ML ๋ชจ๋ธ๋“ค์€, big data๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šตํ•˜๋Š”๋ฐ ์ด๋Ÿฐ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๊ณผ์ •์—์„œ ๋งŽ์€ ๊ฐœ์ธ์ •๋ณด ๋ณด์•ˆ ์ด์Šˆ๊ฐ€ ์ƒ๊ธธ ์ˆ˜ ์žˆ๋‹ค. ์—ฐํ•ฉํ•™์Šต์€, ๋ฐ์ดํ„ฐ๋ฅผ ๋ฉ”์ธ ์„œ๋ฒ„๊ฐ€ ์•„๋‹Œ, ๊ฐœ๊ฐœ์ธ์˜ ๋กœ์ปฌ ํด๋ผ์ด์–ธํŠธ์— ๋‘๊ณ  ๊ทธ ๋กœ์ปฌ ํด๋ผ์ด์–ธํŠธ์—์„œ ํ•™์Šต์„ ์ˆ˜ํ–‰ํ•œ ํ›„, ์—…๋ฐ์ดํŠธ๋œ ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋“ค์„ ์ค‘์•™ ์„œ๋ฒ„๋กœ ๋ณด๋‚ด ์ทจํ•ฉํ•ด์„œ ํ•˜๋‚˜์˜ ๋ชจ๋ธ์„ ์—…๋ฐ์ดํŠธ ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. ๊ทธ๋Ÿฌ๋ฉด ์ค‘์•™ ์„œ๋ฒ„๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์ง€์ง€ ์•Š๊ณ  ์žˆ์Œ์—๋„, ๊ทธ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•ด ํ•™์Šตํ•œ ํšจ๊ณผ๋ฅผ ๋‚ธ๋‹ค.

     

     

    ์—ฐํ•ฉ ํ•™์Šต์˜ ์žฅ์ 

    ๋ฐ์ดํ„ฐ ํ”„๋ผ์ด๋ฒ„์‹œ ํ–ฅ์ƒ (๋ฐ์ดํ„ฐ๊ฐ€ ๋„คํŠธ์›Œํฌ๋ฅผ ํƒ€๊ณ  ์ด๋™ํ•˜์ง€ ์•Š์•„ ์œ ์ถœ ์œ„ํ—˜ ์ค„์–ด๋“ฆ)

    ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ํšจ์œจ์„ฑ(๋„คํŠธ์›Œํฌ ๋น„์šฉ : ๋ฐ์ดํ„ฐ ์ด๋™ > ํŒŒ๋ผ๋ฏธํ„ฐ ์ด๋™)

     

     

    ๊ฐ„๋‹จ ์•Œ๊ณ ๋ฆฌ์ฆ˜

    https://arxiv.org/pdf/1602.05629.pdf

    Federated Averaging Algorithm, McMahan et al. 2016

    1. ์„œ๋ฒ„์— ๋ฉ”์ธ ๋ชจ๋ธ์„ ์ค€๋น„

    2.  K๊ฐœ์˜ ํด๋ผ์ด์–ธํŠธ๋“ค ์ค‘, (ํšจ์œจ์„ฑ์„ ์œ„ํ•ด) ๋žœ๋คํ•˜๊ฒŒ m๊ฐœ๋ฅผ ๊ณ ๋ฅด๊ณ  ์„œ๋ฒ„์— ์žˆ๋Š” ๋ฉ”์ธ ๋ชจ๋ธ์„ ๋ณด๋ƒ„.

    3. ๊ฐ ํด๋ผ์ด์–ธํŠธ ๋‚ด์—์„œ ํด๋ผ์ด์–ธํŠธ๋“ค์ด ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋กœ์ปฌ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•ด์„œ ๋ฉ”์ธ ๋ชจ๋ธ์„ ํ•™์Šต.

    4. ๊ทธ ๊ฒฐ๊ณผ weight๋ฅผ ๋‹ค์‹œ ์„œ๋ฒ„๋กœ ๋ณด๋ƒ„

    5. ๊ทธ๋ ‡๊ฒŒ ๋ชจ์ธ weight๋“ค์˜ ๊ฐ€์ค‘ํ‰๊ท ์„ ์ทจํ•ด ์„œ๋ฒ„์˜ ๋ชจ๋ธ์„ ์—…๋ฐ์ดํŠธ. (๊ฐ€์ค‘ํ‰๊ท  : ๊ฐ ํด๋ผ์ด์–ธํŠธ๊ฐ€ ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ๋ฐ์ดํ„ฐ์ˆ˜์— ๋”ฐ๋ฆ„)

     

    ๊ฐ€์ค‘์น˜ ๊ณ„์‚ฐ ๋ฐฉ๋ฒ•

     

    ์ด๋•Œ, ์ •๋ง ๋‚˜๋ˆ ์ง„ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•ฉ์นœ๋‹ค๊ณ  ์ž˜ ์ž‘๋™ํ• ๊นŒ? ๋ผ๋Š” ์˜๋ฌธ์ ์ด ์ƒ๊น€. ์•„๋ž˜ ์‹คํ—˜์ด ์ด ๊ถ๊ธˆ์ฆ์„ ํ•ด๊ฒฐ.

    ์ด๋Š” MNIST data๋ฅผ federated learningํ•œ loss ๊ฒฐ๊ณผ์ธ๋ฐ, ์™ผ์ชฝ์€ ์ดˆ๊ธฐ์— ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ๊ฐ initializeํ•œ ๋’ค, ๊ฐ์ž ํ›ˆ๋ จ์‹œ์ผœ ํ•ฉ์นœ ๊ฒฐ๊ณผ์ด๊ณ  ์˜ค๋ฅธ์ชฝ์€ ์ดˆ๊ธฐ์— ํ•˜๋‚˜์˜ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ initializeํ•œ ๋’ค, ๊ทธ ๋ชจ๋ธ์— ๋Œ€ํ•ด ๊ฐ๊ฐ ํ›ˆ๋ จ์‹œ์ผœ ํ•ฉ์นœ ๊ฒฐ๊ณผ. ๋™์ผํ•œ ์ดˆ๊ธฐ ๋žœ๋ค ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ๊ฐ๊ฐ ํ•™์Šตํ•œ ๋ชจ๋ธ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•ฉ์นœ ๋ชจ๋ธ์€ ์„ฑ๋Šฅ์ด ํ›จ์”ฌ ์ž˜ ๋‚˜์˜จ๋‹ค.

     

    ๊ทธ๋ฆผ์œผ๋กœ ๋ณด๋Š” ๋™์ž‘ ๋ฐฉ์‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค.

    From federated learning to federated neural architecture search: a survey

     

    ์—ฐํ•ฉํ•™์Šต์˜ ๋‘๊ฐ€์ง€ ๋ฐฉ์‹

    Cross-silo FL

    ๊ธฐ์—…&๋‹จ์ฒด๊ฐ€ ์—ฐํ•ฉ. ์ฆ‰, ํด๋ผ์ด์–ธํŠธ๋Š” ์กฐ์ง์œผ๋กœ ๊ตฌ์„ฑ๋จ. ๊ทธ๋ž˜์„œ ํด๋ผ์ด์–ธํŠธ๋Š” ๋ชจ๋‘ ๋น„์ต๋ช….

    ๋ฏธ๋ฆฌ ๊ณ„ํš๋œ ํ•™์Šต์ด๊ธฐ ๋•Œ๋ฌธ์— ํ•™์Šต์ด๋‚˜ ํ•ฉ์น˜๋Š”๋ฐ ์‹คํŒจํ•  ๊ฐ€๋Šฅ์„ฑ์ด ์ ๋‹ค.

     

    Cross-device FL

    ๋กœ์ปฌ ๋””๋ฐ”์ด์Šค๊ฐ€ ์—ฐํ•ฉ. ์ฆ‰, ํด๋ผ์ด์–ธํŠธ๋Š” ๋งŽ์€ ์ˆ˜์˜ ๋ชจ๋ฐ”์ผ ๊ธฐ๊ธฐ (ํ˜น์€ IoT ๊ธฐ๊ธฐ)

    ๊ฐ ํด๋ผ์ด์–ธํŠธ๊ฐ€ ์ต๋ช…์œผ๋กœ ์กด์žฌํ•˜์—ฌ, "์•…์˜์  ์‚ฌ์šฉ์ž"๊ฐ€ ์กด์žฌํ•  ์ˆ˜ ์žˆ์Œ.

     

     

    ์ง€๊ธˆ๊นŒ์ง€ ๋ฐœ๊ฒฌ๋œ ๋ฌธ์ œ์™€ ํ•ด๊ฒฐ์ฑ…๋“ค

    - ์—ฐํ•ฉํ•™์Šต์„ ์œ„ํ•œ ์ ์ ˆํ•œ ํ›ˆ๋ จ๋ฒ•

    ๊ธฐ์กด์—๋Š” ๋ณดํ†ต ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ์…‹์— ์ ํ•ฉํ•œ SGD๊ฐ™์€ ์ตœ์ ํ™” ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉ. ๊ทธ๋Ÿฌ๋‚˜ ์—ฐํ•ฉ ํ•™์Šต์—์„œ ๋ฐ์ดํ„ฐ๋Š” ๋งค์šฐ ๊ณ ๋ฅด์ง€ ์•Š์€ ๋ฐฉ์‹์œผ๋กœ ์ˆ˜๋ฐฑ๋งŒ ๋Œ€์˜ ์žฅ์น˜์— ๋ถ„์‚ฐ๋˜์–ด์žˆ์Œ. ๋˜ํ•œ ์ด๋Ÿฌํ•œ ์žฅ์น˜(๋กœ์ปฌ ๋””๋ฐ”์ด์Šค)๋Š” ๋Œ€๊ธฐ ์‹œ๊ฐ„์ด ํ›จ์”ฌ ๋” ๊ธธ๊ณ  ์ฒ˜๋ฆฌ๋Ÿ‰์ด ๋‚ฎ์€ ์—ฐ๊ฒฐ์„ ์ œ๊ณตํ•˜๋ฉฐ ๊ฐ„ํ—์ ์œผ๋กœ๋งŒ train์— ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

    => ์ตœ์‹  ๋ชจ๋ฐ”์ผ ์žฅ์น˜์˜ ๊ฐ•๋ ฅํ•œ ํ”„๋กœ์„ธ์„œ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋‹จ์ˆœํ•œ ๊ทธ๋ž˜๋””์–ธํŠธ ๋‹จ๊ณ„๋ณด๋‹ค ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ์—…๋ฐ์ดํŠธ๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ

     

    - ์ •๋ง ์•ˆ์ „ํ•œ๊ฐ€? (๋ฐ์ดํ„ฐ ์œ ์ถœ ๋ฌธ์ œ)

    Deep Leakage from Gradients : gradient๋ฅผ ๊ฐ€์ง€๊ณ  ์—ญ์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ๋ณต์›ํ•จ. (dummy ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“  ๋’ค ์ด ์ž…๋ ฅ์— ๋Œ€ํ•œ ํ˜„์žฌ ๋ชจ๋ธ์˜ gradient๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ , ์ด gradient๊ฐ€ ๋ชฉํ‘œํ•˜๋Š” gradient์— ๊ฐ€๊นŒ์›Œ์ง€๋„๋ก ๋ฐ์ดํ„ฐ ์ž์ฒด๋ฅผ update)

     

    - ๋ชจ๋ธ์— ๋‹ค๋ฅธ ์˜ํ–ฅ์€์—†๋Š”๊ฐ€?

    Adversarial attack(poisoning attack) : local model๋“ค์„ ํ†ตํ•ฉํ•ด์„œ global model์„ ์—…๋ฐ์ดํŠธํ•˜๋Š” ๋ฐฉ์‹์ด๋ผ ํด๋ผ์ด์–ธํŠธ๊ฐ€ ์ž์‹ ์˜ local model์— ๋ชจ๋ธ์— ๋Œ€ํ•ด์„œ adversarial ๋ฐ์ดํ„ฐ๋ฅผ ๋„ฃ์œผ๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ๋‚ฎ์•„์ง.

    =>  ๋ฐฉ์–ด๊ธฐ๋ฒ•(adversarial training)๋„ ์—ฐ๊ตฌ์ค‘

     

    (์—ฌ๊ธฐ์„œ๋ถ€ํ„ฐ๋Š” https://arxiv.org/abs/1912.04977 ์ฐธ์กฐ)

    - ์ค‘์•™์„œ๋ฒ„์˜ ๋ฌธ์ œ์ 

    ๋‹จ์ผ ์„œ๋ฒ„ ๋ชจ๋ธ์ด๋ผ single point of failure ๋ฌธ์ œ๊ฐ€ ์žˆ๊ณ , ํ•ญ์ƒ ์„œ๋น„์Šค ๊ฐ€๋Šฅํ•œ ์ƒํƒœ๊ฐ€ ์•„๋‹ˆ๋ผ ์žฅ๊ธฐ์ ์œผ๋กœ ๋ณด๋ฉด ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜๋„ ์žˆ์Œ

    => ํ•ด๊ฒฐ์ฑ… : Fully Decentralized Learning (์™„์ „ ํƒˆ์ค‘์•™ ํ•™์Šต) : ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ๋ฐฉ์‹์„ Peer-to-Peer(P2P)๋กœ ๋ณ€ํ™”

     

    - P2P ๋ฐฉ์‹์˜ ์™„์ „ ํƒˆ์ค‘์•™ ํ•™์Šต์˜ ๋ฌธ์ œ์ 

    1. ๋„คํŠธ์›Œํฌ ํ† ํด๋กœ์ง€&๋น„๋™๊ธฐ ํ†ต์‹  ๋ฌธ์ œ

    : ๋ชจ๋“  ๋…ธ๋“œ๋Š” ์™„์ „ํ•˜๊ฒŒ ์—ฐ๊ฒฐ๋˜์–ด ์žˆ์ง€๋งŒ ๋ฉ”์‹œ์ง€๊ฐ€ ํŠน์ • ํ™•๋ฅ ๋กœ ๋„๋‹ฌํ•˜์ง€ ์•Š๋Š” ์ƒํ™ฉ์—์„œ๋„ ๊ฒฌ๊ณ ํ•จ(robustness)์ด ํ•„์š”.

     

    2. ํƒˆ์ค‘์•™ SGD์˜ ๋กœ์ปฌ ์—…๋ฐ์ดํŠธ ๋ฌธ์ œ
    : ๋กœ์ปฌ ๋””๋ฐ”์ด์Šค๊ฐ€ ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ๊ณผ ์ปค๋ฎค๋‹ˆ์ผ€์ด์…˜ ํ•˜๊ธฐ ์ „์— ๋ช‡ ๋ฒˆ์˜ ๋กœ์ปฌ ์—…๋ฐ์ดํŠธ๋ฅผ ๋ฐœ์ƒํ•  ๊ฒƒ์ธ์ง€ ๊ฒฐ์ •ํ•˜๋Š” ๊ฑด ์ค‘์š”ํ•œ ๋ฌธ์ œ๋‹ค.(why?)

    ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ๋ฒˆ์˜ ๋กœ์ปฌ ์—…๋ฐ์ดํŠธ ์ดํ›„ ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ๋กœ ํ•ฉ์น˜๋Š”๊ฒŒ Non-IID ๋ฐ์ดํ„ฐ ํ™˜๊ฒฝ์—์„œ ์ˆ˜๋ ด์ด ์ž˜ ๋œ๋‹ค๊ณ  ์•Œ๋ ค์ ธ ์žˆ์ง€๋งŒ, ํ†ต์‹  ๋น„์šฉ์ด ์ฆ๊ฐ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋น ๋ฅด๊ฒŒ ์ˆ˜๋ ด์„ ํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ต๋‹ค.

     

    3. ์‹ ๋ขฐ ๋ฌธ์ œ

    P2P๋Š” ์ฐธ์—ฌ์™€ ์ดํƒˆ์ด ์ž์œ ๋กœ์›Œ์„œ ์•…์˜์ •์ธ ๊ณต๊ฒฉ์ž๋‚˜ ์‹ ๋ขฐํ•  ์ˆ˜ ์—†๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ํ•™์Šต์— ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์Œ.

     

    4. ๊ฐœ์ธํ™”

    : ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ์„ ๊ฐ ๋กœ์ปฌ ๋””๋ฐ”์ด์Šค์˜ ์šฉ๋„์— ๋”ฐ๋ผ ๋ณ€ํ˜•ํ•˜์—ฌ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์–ด์•ผ ํ•œ๋‹ค๋Š” ํŠน์ง•. (Ex/ next-word prediction ์–ธ์–ด ๋ชจ๋ธ์„ ์ƒ์„ฑํ•œ๋‹ค๊ณ  ํ•  ๋•Œ, ๊ฐ ๋กœ์ปฌ ๋””๋ฐ”์ด์Šค๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์–ธ์–ด๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ์„ ๋กœ์ปฌ์— ์ ์šฉํ•  ๋•Œ ๋‹ค๋ฅธ feature๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์ด ํ•„์š”) => ๊ฐ„๋‹จํ•˜๊ฒŒ๋Š” User Context ๊ฐ’์„ ์ž…๋ ฅ๊ฐ’์— ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•. + ์ตœ๊ทผ์—๋Š” ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ์„ ๋กœ์ปฌ์—์„œ Fine-Tuningํ•˜๋Š” ์—ฐ๊ตฌ ์ง„ํ–‰ ์ค‘.

     

    => ํ•ด๋‹น ๋…ผ๋ฌธ์—์„œ๋Š” ์™„์ „ ํƒˆ์ค‘์•™ ํ•™์Šต์˜ ๋ฌธ์ œ ํ•ด๊ฒฐ์ฑ…์œผ๋กœ ๋ธ”๋ก์ฒด์ธ๊ณผ ์Šค๋งˆํŠธ ์ปจํŠธ๋ž™ํŠธ๋ฅผ ์ž ์žฌ์  ํ•ด๊ฒฐ์ฑ…์˜ ํ›„๋ณด๊ตฐ์œผ๋กœ ์–ธ๊ธ‰

     

    - ๋ฐ์ดํ„ฐ์˜ non-IID ๋ฌธ์ œ

    Non-IID(Independent and Identically Distributed) : ๋ฐ์ดํ„ฐ๊ฐ€ "๋น„๋…๋ฆฝ์ "์ด๊ณ , "๋™์ผํ•˜์ง€ ์•Š๊ฒŒ ๋ถ„์‚ฐ"๋จ.

    ์ฆ‰, ๊ฐ ํด๋ผ์ด์–ธํŠธ๋“ธ ๋ฐ์ดํ„ฐ๋Š” ์ข…์†์ ์ผ ์ˆ˜ ์žˆ๊ณ , ๋™์ผํ•œ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์ง€ ์•Š๋‹ค๋Š” ๋œป์ด๋‹ค. ๋•Œ๋ฌธ์— ๊ธ€๋กœ๋ฒŒํ•˜๊ฒŒ ์ตœ์ ํ™”๋œ ๋ชจ๋ธ์„ ๋งŒ๋“œ๋Š” ๋ฐ ์–ด๋ ค์›€์ด ์žˆ๋‹ค. 

    => ์ตœ๊ทผ์— ๋‚˜์˜จ ํ•ด๊ฒฐ์ฑ… : ๋ฐ์ดํ„ฐ์…‹ ์ฆ๋ฅ˜(dataset distillation)

     

    - Adversarial attack ๋ฌธ์ œ

    ๋ณดํ†ต ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ์ค‘์•™ ์„œ๋ฒ„์—์„œ๋งŒ ์กด์žฌํ•ด, ํด๋ผ์ด์–ธํŠธ๋Š” ๋ชจ๋ธ ๋‚ด๋ถ€๋ฅผ ๋ณผ ์ˆ˜ ์—†๋Š” "black-box" ์‹œ์Šคํ…œ์ด๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ FL์€ ๋กœ์ปฌ ํด๋ผ์ด์–ธํŠธ์—๊ฒŒ ๋ฉ”์ธ ๋ชจ๋ธ์ด ์ „๋‹ฌ๋˜๊ณ , ๊ฑฐ๊ธฐ์„œ ์—…๋ฐ์ดํŠธ ํ•˜๋‹ค ๋ณด๋‹ˆ ์‚ฌ์šฉ์ž๊ฐ€ ๋ชจ๋ธ์˜ ๋‚ด๋ถ€๋ฅผ ๋ณผ ์ˆ˜ ์žˆ๋Š” "white-box" ์‹œ์Šคํ…œ์ด๋‹ค. ์ƒ๋Œ€์ ์œผ๋กœ black box๋ชจ๋‹ค white box๊ฐ€ ์ ๋Œ€์ ์ธ ๊ณต๊ฒฉ์— ํ›จ์”ฌ ์ทจ์•ฝํ•˜๋ฉฐ, ์•„์ง ์—ฐ๊ตฌ๋œ ๋ฐ”๋Š” ์—†์ง€๋งŒ FL์—์„œ๋งŒ ํ†ตํ•˜๋Š” ์ž ์žฌ์ ์ธ adversarial vector๊ฐ€ ์žˆ์„ ์ˆ˜๋„ ์žˆ๋‹ค. 

    ๊ณต๊ฒฉ ๋ฐฉ์‹์—๋Š” Model update poisoning (ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ’ ์˜ค์—ผ), data poisoning(๋ฐ์ดํ„ฐ ์˜ค์—ผ), evasion attack(adversarial examples)

     

     

     

    ์‘์šฉ ๋ถ„์•ผ

    - Intelligent recommendation

    1. ๊ตฌ๊ธ€ ํ‚ค๋ณด๋“œ์˜ ์ถ”์ฒœ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ๊ฐœ์„  : Gboard

    gboard์— ์ œ์•ˆ๋œ ์ฟผ๋ฆฌ๊ฐ€ ํ‘œ์‹œ๋˜๋ฉด, ํœด๋Œ€์ „ํ™”๋Š” ๊ทธ ์ œ์•ˆ์„ ํด๋ฆญํ–ˆ๋Š”์ง€ ์—ฌ๋ถ€์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ๋กœ์ปฌ์— ์ €์žฅ. Federated learning์€ ๊ธฐ๊ธฐ์—์„œ ๊ธฐ๋ก์„ ์ฒ˜๋ฆฌํ•ด์„œ gboard์˜ ์ฟผ๋ฆฌ์ œ์•ˆ๋ชจ๋ธ์˜ ๋‹ค์Œ ๋ฐ˜๋ณต์— ๋Œ€ํ•œ ๊ฐœ์„  ์‚ฌํ•ญ์„ ์ œ์•ˆ.

    (๋กœ์ปฌ ๋””๋ฐ”์ด์Šค ๋‚ด Train์—์„œ๋Š” TensorFlow ์˜ ์ถ•์†Œ ๋ฒ„์ „ ์‚ฌ์šฉ + ์žฅ์น˜๊ฐ€ ์œ ํœด ์ƒํƒœ + ์ถฉ์ „์ค‘ + ์™€์ดํŒŒ์ด ์ผ ๋•Œ๋งŒ train ์ˆ˜ํ–‰ => ํœด๋Œ€ํฐ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š๋Š”๋‹ค(์ฃผ์žฅ))

    https://ai.googleblog.com/2017/04/federated-learning-collaborative.html

     

    Federated Learning: Collaborative Machine Learning without Centralized Training Data

    Posted by Brendan McMahan and Daniel Ramage, Research Scientists Standard machine learning approaches require centralizing the training data...

    ai.googleblog.com

     

    2. ๋” ์ ์€ ๋ฌธ์ž๋ฅผ ์ž…๋ ฅํ•ด์„œ ์›ํ•˜๋Š” ์›น์‚ฌ์ดํŠธ ์ฐพ์„ ์ˆ˜ ์žˆ๋„๋ก ๋„์›€.

     

     

    - Vehicular network (์ฐจ๋Ÿ‰ ๋„คํŠธ์›Œํฌ)

    GPS์—์„œ ๊ฐ์ง€ํ•œ ์œ„์น˜ ๋ฐ ๋ฐฉํ–ฅ, ์ฐจ๋Ÿ‰์— ํƒ‘์žฌ๋œ ์นด๋ฉ”๋ผ๋กœ ์บก์ฒ˜ํ•œ ์ด๋ฏธ์ง€, ์œ ์•• ์„ผ์„œ์˜ ์••๋ ฅ ๋ฐ์ดํ„ฐ์™€ ๊ฐ™์ด ์ฐจ๋Ÿ‰์—์„œ ์žฅ์น˜๊ฐ€ ์ƒ์„ฑํ•˜๋Š” ๋ฐ์ดํ„ฐ๋Š” ์ž๋™์ฐจ ์ œ์กฐ์—…์ฒด๊ฐ€ ์ง€๋Šฅํ˜• ๋‚ด๋น„๊ฒŒ์ด์…˜ ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์กฐ๊ธฐ์— ์กฐ๊ธฐ์— ๋ฐœ๊ฒฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ท€์ค‘ํ•œ ์ž์› => ์ด๋ฅผ ์ด์šฉํ•ด์„œ train์„ ํ•  ์ˆ˜ ์žˆ์Œ.

     

    - Healthcare system

    1. ์˜๋ฃŒ ๋ถ„์•ผ : ํ™˜์ž ๋ฐ์ดํ„ฐ์™€ ๊ฐ™์ด ๋ฏผ๊ฐํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์œ ์ถœํ•˜์ง€ ์•Š์œผ๋ฉด์„œ ๋ชจ๋ธ์„ ํ•™์Šต ๊ฐ€๋Šฅ.

     

    ๊ทธ ์™ธ : A Survey on Federated Learning and its Applications for Accelerating Industrial Internet of Things

    (https://arxiv.org/ftp/arxiv/papers/2104/2104.10501.pdf

    - ํ˜„์žฌ ์‘์šฉ(์—ฐ๊ตฌ) ๋ถ„์•ผ

    FL for IIoT

    Healthcare & Medical (HM)

    Recommender System (RS)

    Smart Transportation (ST)

    Localization Service (LS)

    Mobile Packet Classification (MPC)

    Payment in Smart Finance (SF)

    Data Relevance Analysis (DRA)

    Object Detection in Computer Vision (CV)

    Traffic Sign Classification in Smart City (SC)

    Energy Prediction in Smart Energy (SE)

    Anomaly Detection and Voice Assistant in Smart Building (SB)

    Collision Detection and Imitation Learning in Autonomous Driving (AD)

     

     

    - ์•ž์œผ๋กœ ๊ฐ€๋Šฅ์„ฑ ์žˆ๋Š” ์‘์šฉ ๋ถ„์•ผ

    AR/VR-guided operations

    Precise robotics collaboration

    Industrial environmental monitoring.

    Product defect detection.

    Optimal supply chain scheduling

    Generative product design

    Security

     

     

     

     


    reference

    https://seewoo5.tistory.com/22

     

    Privacy Preserving Machine Learning (1) - Federated Learning

     ๋จธ์‹ ๋Ÿฌ๋‹๊ณผ ๋”ฅ๋Ÿฌ๋‹์ด ๊ฐ๊ด‘์„ ๋ฐ›์œผ๋ฉด์„œ ๊ทธ์— ๋Œ€ํ•œ ์ˆ˜์š” ์—ญ์‹œ ๊ธ‰์ฆํ•˜๊ฒŒ ๋˜์—ˆ๋Š”๋ฐ์š”, (๋ฐฉ๋Œ€ํ•œ) ๋ฐ์ดํ„ฐ์— ๊ธฐ๋ฐ˜์„ ๋‘๊ณ  ์žˆ๋Š” ML/DL ๊ธฐ์ˆ ๋“ค์„ ์ด์šฉํ•จ์— ์žˆ์–ด์„œ ๋ฐ์ดํ„ฐ์˜ ๋ณด์•ˆ์— ๋Œ€ํ•ด์„œ ํ•œ๋ฒˆ์ฏค์€ ์ƒ๊ฐ

    seewoo5.tistory.com

    https://zzaebok.github.io/federated_learning/machine_learning/Federated_Learning/

     

    Federated Learning ์ •๋ฆฌ

     

    zzaebok.github.io

    https://medium.com/curg/%EC%97%B0%ED%95%A9-%ED%95%99%EC%8A%B5-federated-learning-%EA%B7%B8%EB%A6%AC%EA%B3%A0-%EC%B1%8C%EB%A6%B0%EC%A7%80-b5c481bd94b7

     

    ์—ฐํ•ฉ ํ•™์Šต(Federated Learning), ๊ทธ๋ฆฌ๊ณ  ์ฑŒ๋ฆฐ์ง€

    ์—ฐํ•ฉ ํ•™์Šต(FL: Federated Learning)์€ ๋‹ค์ˆ˜์˜ ๋กœ์ปฌ ํด๋ผ์ด์–ธํŠธ์™€ ํ•˜๋‚˜์˜ ์ค‘์•™ ์„œ๋ฒ„๊ฐ€ ํ˜‘๋ ฅํ•˜์—ฌ ๋ฐ์ดํ„ฐ๊ฐ€ ํƒˆ์ค‘์•™ํ™”๋œ ์ƒํ™ฉ์—์„œ ๊ธ€๋กœ๋ฒŒ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๋Š” ๊ธฐ์ˆ ์ด๋‹ค. ์—ฌ๊ธฐ์„œ ๋กœ์ปฌ ํด๋ผ์ด์–ธํŠธ๋Š” ์‚ฌ๋ฌผ ์ธํ„ฐ

    medium.com

    https://www.hitechnectar.com/blogs/applications-of-federated-learning/

     

    5 Applications of Federated Learning

    Federated learning applications is a booming technique. Its applications are spread over a number of industries including Healthcare Industry, Insurance Sector, IoT.

    www.hitechnectar.com

    https://arxiv.org/ftp/arxiv/papers/2104/2104.10501.pdf

     

    ๋Œ“๊ธ€