Upload filter

from Wikipedia, the free encyclopedia

An upload filter is a server-side software that checks files when they are uploaded , rejects them if necessary, changes them or initiates other measures. Machine data processing is intended to replace a content check by humans . Such filters are important for publicly available services if they make content of other users accessible to third parties or publish it in full, for example social media or video portals . Upload filters can prevent legal interests such as copyright or personal rights from being violated. Other filters only allow certain types of files or limit the size of a file. They also serve to rationalize work processes and can protect workers who perform manual inspections from psychological stress.

Techniques of automated image , speech and text recognition are used for upload filters . This means that copies are easily recognizable, but the reliability of the recognition is sometimes controversial for more complex requirements. Critics warn of the danger of overblocking if exceptions and special cases cannot be reliably identified as such.

Use and technology

When uploading files, for example, file properties such as file format and file size can be checked. A computer program blocks files that violate filter rules. In the case of content restrictions, the implementation in program code is more complex. An implementation could look like this: Certain features of the content of the file are worked out and based on these features it is decided whether the file is still within the restrictions and may be loaded. Depending on the complexity of the content and situations to be recognized, the recognition is associated with different error rates, more on this in the section on susceptibility to errors .

Some providers use upload filters to detect copyright infringements. State of the art (2019) for this use case is a simple pattern comparison, whereas artificial intelligence methods have not been used so far. A digital fingerprint (e.g. the acoustic fingerprint for music) is typically used for such pattern comparisons and the fingerprint of newly uploaded content is searched for in a database of stored fingerprints. The Content ID system YouTube uses such upload filter one - but with follow-through thousands employees.

In PhotoDNA is one of Microsoft with the University of Dartmouth developed technology, which the authorities known child pornography recognize material and to reject. To compare recordings, digital fingerprints are created and compared based on features such as color gradients and edges . This technology is used by several providers.

The Facebook group uses upload filters to identify revenge porn . Digital fingerprints are used here, for which those affected transmit the photo to Facebook. The fingerprint is calculated from this. In addition, algorithms from the field of machine learning should also recognize unknown recordings.

Upload filters for filtering extremist content are also being considered; this is also proposed , among other things, in the “EU Regulation on Preventing the Dissemination of Terrorist Online Content” (TERREG).

Legal and political situation

At the beginning of 2012, the European Court of Justice ruled that social networks are not obliged to use automated upload filtering to check user contributions for copyright infringements due to the prohibition of a general monitoring obligation. On the other hand, this affects entrepreneurial freedom, since this requires expensive IT systems. A collecting society had sued the social network Netlog .

For years, attempts have been made to better protect and monetize intellectual property at European level . There are many different legal situations in the EU, for example with regard to freedom of panorama . This led to a certain lack of law on digital platforms and thus to new freedoms. This is intended to limit the directive of the European Parliament and of the Council on copyright in the digital single market .

In the coalition agreement of the 19th parliamentary term of the Bundestag , decided on February 7, 2018, the term upload filter was used for the first time in Germany at the highest political level. The position of the Merkel IV cabinet for the copyright reform of the European Union was thus apparently determined.

“We reject an obligation on platforms to use upload filters to“ filter ”content uploaded by users for copyright infringing content as disproportionate. Negative effects on small and medium-sized publishers must be avoided. "

- Coalition agreement between CDU, CSU and SPD.

As the EU directive passed through the EU Parliament, there were major discussions inside and outside the Parliament. In Germany and elsewhere, hundreds of thousands of people demonstrated against the EU directive on upload filtering.

Due to the complexity of the technology and the large number of possible rights holders, Ulrich Kelber, as the Federal Commissioner for Data Protection and Freedom of Information, warned against the emergence of technical oligopolies , as the high development costs force many data movements to be routed through the filters of fewer providers. According to Internet pioneers Tim Berners-Lee and Vint Cerf , mandatory upload filters of this type could cause great damage to the free Internet.

The EU directive was passed in the European Parliament on March 26, 2019 in the form that the use of upload filters could become the new standard on digital platforms. The member states have two years to formulate the EU directive in national legal texts. France was the first country to present a draft law in December 2019, which places a strong focus on the interests of rights exploiters and leaves out many of the directive's protective provisions for internet users. There are, for example, no exceptions for parodies, caricatures or quotations, that platforms cannot be obliged to generally monitor all uploads, and there is also no prohibition against legally blocking content. The further refinement makes the use of upload filters unavoidable, since rights holders can refuse to grant licenses and platforms have to prove that they have made every effort to conclude license agreements with the rights holders in order to avoid liability.

Internet censorship and restriction of freedom

According to critics, such filters sometimes help to operate Internet censorship and to block content that violates a certain opinion in advance. This can endanger the right to freedom of expression and freedom of the press . This also includes the term machine censorship.

Susceptibility to errors

A comparison of two pieces of content to determine whether one piece of content is a copy of the other works very well. PhotoDNA is ascribed a comparatively high level of reliability.

Algorithms reach their limits when files are changed or manipulated. It becomes difficult when context has to be included. This is not the case in every application. The main area of ​​application of PhotoDNA is, for example, the detection of child pornographic material. This is illegal in any context. When detecting copyright infringements, however, material must be assessed differently depending on the context in order to distinguish permitted parodies and quotations from prohibited copyright infringements. Artificial intelligence methods are still immature. The reliable automatic classification of such cases is therefore currently not considered possible by Florian Gallwitz, professor for media informatics , so that undesirable side effects such as overblocking will inevitably occur in this area of ​​application.

Many web video providers criticize the loss of advertising income due to incorrect classification of their content through upload filters. Videos that deal with critical topics in a factual or entertaining way can be blocked without violating guidelines based on the content and certain schematic features of the filter. It is also very questionable to pull individual words in a video out of context, for example by evaluating a played insult as hate. This restricts artistic freedom, for example in satire .

However, some upload filters should not only recognize known material, but should also be able to classify new content based on machine learning, such as the upload filters used by Facebook. However, this always leads to incorrect classifications. In order to counteract false identifications to be expected with this approach, Facebook also subjects recordings to the manual check.

When terrorist content was detected on the YouTube video platform, according to the provider, one million videos were classified as suspicious by its own filter software in the first quarter of 2019. 90,000 of these videos were confirmed as correct in a manual check, which corresponds to an accuracy of less than 10%.

See also

Web links

Individual evidence

  1. ↑ a b c d so Frank Pallas, TU Berlin, on Deutschlandfunk: Jan Rähm: Uploadfilter. Why Critics Are Afraid of Censorship. March 9, 2019, accessed on March 22, 2019 (many technical details plus a brief history).
  2. ↑ a b Florian Gallwitz, professor for media informatics, Technical University of Nuremberg Georg Simon Ohm: EU copyright reform: experts on upload filters. On the question of how reliable upload filters would be in difficult cases, such as when assessing satirical contributions. March 8, 2019, accessed on March 13, 2019 (received by Deutsche Welle : Memes could be filtered out by EU copyright law , March 13, 2019).
  3. ↑ PhotoDNA: step-by-step. Microsoft, September 21, 2013, archived from the original ; accessed on March 14, 2019 (English).
  4. ↑ a b heise online: Upload filter: Facebook and Instagram automatically delete revenge porn. Retrieved March 17, 2019 .
  5. ↑ Comment by Simon Hurtz: This compromise endangers the free network . In: sueddeutsche.de . February 13, 2019, ISSN  0174-4917 ( sueddeutsche.de [accessed on March 13, 2019]).
  6. ↑ Sonja Peteranderl, DER SPIEGEL: This is how the EU wants to contain terror content on the Internet - DER SPIEGEL - Netzwelt. Retrieved January 14, 2020 .
  7. ↑ ECJ stops upload filters for social networks. In: sueddeutsche.de. February 16, 2012, accessed January 4, 2019 .
  8. ↑ A new departure for Europe. A new dynamic for Germany. A new cohesion for our country. Coalition agreement between the CDU, CSU and SPD. 19th legislative period (Merkel IV cabinet) , page 49 on lines 2212–2214, March 12, 2018, accessed on March 7, 2019.
  9. ↑ Federal Commissioner for Data Protection and Freedom of Information, reform of copyright law also harbors data protection risks , Bonn / Berlin February 26, 2019 (press release online)
  10. ↑ Open Letter of the EFF: Article 13 of the EU Copyright Directive Threatens the Internet , June 12, 2018
  11. ↑ N ° 2488 - Projet de loi relatif à la communication audiovisuelle et à la souveraineté culturelle à l'ère numérique. Retrieved January 14, 2020 (French).
  12. ↑ Julia Reda: France presents upload filter law, “forgets” user rights. In: JuliaReda.eu. Retrieved January 14, 2020 .
  13. ↑ Holger Bleich: EU Commission calls for the use of upload filters. In: heise.de. Retrieved March 17, 2018 .
  14. ↑ a b Interview with Simon Hurtz: These upload filters would be real censorship machines . In: sueddeutsche.de . June 20, 2018, ISSN  0174-4917 ( sueddeutsche.de [accessed March 14, 2019]).
  15. ↑ YouTube - MrTrashpack criticizes badly slumped revenues. In: playnation.de. Retrieved March 17, 2018 .
  16. ↑ Benjamin Köhler: YouTube annoys users with advertising guidelines. In: derwesten.de. Retrieved March 17, 2018 .
  17. ^ David Shepardson: Google spends hundreds of millions of dollars on content review: letter. Reuters, May 2, 2019, accessed May 6, 2019 .