[{"data":1,"prerenderedAt":142},["ShallowReactive",2],{"project-28105849":3},{"id":4,"name":5,"line":5,"line_source":6,"title":7,"url":8,"domain":9,"maker":10,"posted_at":11,"year":12,"points":13,"comments":14,"rank_in_year":15,"shown_in_year":16,"top":17,"tier":18,"category":19,"subcategory":20,"format":21,"keywords":22,"reception":27,"sentiment":28,"stance_mix":29,"scored_comments":32,"aspects":33,"strip":36,"post_text":28,"what":38,"said":57,"maker_in_thread":85,"shape":92,"grid":95,"attempts":106,"same_week":107,"maker_shows":93,"maker_hidden":135,"reader_version":136,"card":137},"28105849","Generating Collisions on NeuralHash","title","Show HN: Generating Collisions on NeuralHash","https:\u002F\u002Fgist.github.com\u002Funrealwill\u002Fc480371c3a4bf3abb29856c29197c0be","gist.github.com","GistNoesis",1628420752,2021,22,6,1321,2533,53,"shown","security_privacy","perceptual hash attacks","other",[23,24,25,26],"perceptual hash","adversarial attack","neuralhash","hash collision","quiet",null,{"positive":30,"neutral":31,"mixed":30,"critical":31},0,1,2,{"docs":34,"idea":35},{"positive":30,"critical":31},{"positive":30,"critical":30},[30,31,30,37,30,30],3,{"problem":39,"audience":28,"pricing":28,"motivation":49,"time_to_build":28,"stack":54,"openness":55,"stage":56,"team_size":55},{"value":40,"source":41,"evidence":42},"Perceptual hashes like NeuralHash can be fooled into producing chosen collisions via gradient-based optimisation","maker",[43,46],{"id":44,"by":10,"text":45},"28105858","This is a proof of concept on generating collisions on a perceptual hash similar to the one Apple recently decided to use to scan your devices for CSAM.…",{"id":47,"by":10,"text":48},"28106158","It's a shorter version of \"Generating Collisions on a Neural Perceptual Hash like Apple's NeuralHash\" optimized for click-baitingness ;), but the spirit of the attack is there.…",{"value":50,"source":41,"evidence":51},"Show that Apple's planned CSAM scanning hash is vulnerable to adversarial collisions",[52,53],{"id":44,"by":10,"text":45},{"id":47,"by":10,"text":48},[],"unknown","prototype",{"praise":58,"objections":67,"requests":75,"questions":76,"alternatives":84},[59],{"value":60,"source":61,"evidence":62},"Commenter says the hash-tolerance property explains why perceptual hashes are poor evidence","commenter",[63],{"id":64,"by":65,"text":66},"28115783","chaboud","What you're highlighting is the lack of avalanche effect with perceptual hashes (on purpose). This makes them great for identifying transformed content (e.g. local TV channels…",[68],{"value":69,"source":61,"evidence":70},"Title overstates the work since the maker lacks Apple's actual architecture and weights",[71],{"id":72,"by":73,"text":74},"28105896","krapp",">I don't have the architecture nor the weights of the actual neural hash they use so this is only a POC in the general case. Then why is this titled \"Generating Collisions on…",[],[77],{"value":78,"source":61,"evidence":79},"How steganographic content would affect the perceptual hash similarity",[80],{"id":81,"by":82,"text":83},"28113167","tehnub","In the manual review failure case, you're saying that the reviewer would decrypt the steganographical encryption, discover the hidden offensive content, and thereby decide that…",[],{"comments":37,"answered":32,"key":86},[87,88,89],{"id":44,"text":45},{"id":47,"text":48},{"id":90,"text":91},"28113441","My guess is that adding steganographically hidden content is almost free for the attacker as perceptual hash is resistant to small variation. There is a free parameter called gap…",{"commenters":93,"max_depth":93,"maker_replies":37,"first_comment_mins":32,"lifespan_hours":94},4,25,[96,98,100,102,103,104],[44,30,30,97,10,31],"project\u002Fneutral\u002Fidea",[72,30,31,99,73,30],"maker\u002Fcritical\u002Fdocs",[47,30,30,101,10,31],"maker\u002Fneutral\u002Fdocs",[81,30,37,97,82,30],[90,30,30,97,10,31],[64,30,30,105,65,30],"commenter\u002Fneutral\u002Fidea",[],[108,113,118,123,127,131],{"id":109,"name":110,"points":111,"category":112},"28074827","Repo Visualization",283,"dev_tools",{"id":114,"name":115,"points":116,"category":117},"28088882","GrowthBook",199,"data",{"id":119,"name":120,"points":121,"category":122},"28120247","colab.research.google.com",169,"ai",{"id":124,"name":125,"points":126,"category":112},"28084104","Systemizer",167,{"id":128,"name":129,"points":130,"category":122},"28100346","Stray Robots toolkit",166,{"id":132,"name":133,"points":134,"category":112},"28128445","Lona",165,false,"v2",{"id":138,"name":5,"line":5,"ai":135,"cat":19,"year":12,"points":13,"comments":14,"rec":27,"s":28,"top":17,"kw":139,"maker":10,"domain":9,"tier":18,"attempts":30,"strip":140,"url":8,"mix":141,"praise":60,"objection":69,"sub":20},28105849,[23,24,25,26],"010300",[30,31,30,31],1794036609863]